Tilburg University
Empirical studies on the cross-section of corporate bond and stock markets van Zundert, Jeroen
Publication date: 2018
Document Version Publisher's PDF, also known as Version of record
Link to publication in Tilburg University Research Portal
Citation for published version (APA): van Zundert, J. (2018). Empirical studies on the cross-section of corporate bond and stock markets. CentER, Center for Economic Research.
General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal Take down policy If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Download date: 01. okt. 2021 Empirical Studies on the Cross-Section of Corporate Bond and Stock Markets
Jeroen van Zundert
January 2018
Empirical Studies on the Cross-Section of Corporate Bond and Stock Markets
Proefschrift
ter verkrijging van de graad van doctor aan Tilburg University op gezag van de rector magnificus, prof. dr. E.H.L. Aarts, in het openbaar te verdedigen ten overstaan van een door het college voor promoties aangewezen commissie in de aula van de Universiteit op
vrijdag 19 januari 2018 om 14.00 uur door
Jeroen Adrianus Catharina van Zundert geboren op 18 juli 1989 te Zundert. Promotiecommissie:
Promotores: prof. dr. J.J.A.G. Driessen prof. dr. F.C.J.M. de Jong
Overige leden: dr. L.T.M. Baele dr. E. Eiling prof. dr. F.A. de Roon prof. dr. A.C.F. Vorst Acknowledgements
This dissertation marks the end of my PhD track at Tilburg University. Offi- cially, this project started two years ago, but looking back it really has been the result of the support of many and my passion for economics and math- ematics ever since my childhood. As a 12-year old, I was already intrigued by financial markets. After cycling back from school to my parents’ home, I turned on the tv at 16.00, just in time for a new market update on RTL Z, a Dutch financial tv channel. The U.S. stock exchanges had just opened, and that usually meant markets were moving. I did not fully understand why markets were moving, but somehow this routine continued throughout sec- ondary school. At the sixth grade, I had to pick a bachelor study. However, I could not find a study I liked until I found out about econometrics, a study combining mathematics with economics. I was immediately sold, and during the years at Tilburg University I only got more excited about econometrics. During the master’s phase a course called “Empirical Finance” was taught. I remember I was not too excited about this course, as all this research on the efficiency of markets seemed very abstract to me. However, during that course, there was a guest lecture from amongst others Laurens Swinkels, telling about the work he did at Robeco, and, of course, advertising the Robeco Super Quant internship program. That lecture had stuck in my mind, as in my final year I decided to apply at Robeco to write my master thesis. The interview was with Daniel Haesen and Patrick Houweling, and although it felt I would not be hired as I didn’t knew anything about these instruments called “corporate bonds” (I even failed to mention interest rate risk as being important for corporate bonds. . . ), they still decided to hire me, for which I am very grateful. Thanks to their dedication throughout the internship, I could quickly get my (corporate bonds) knowledge up to speed and write my thesis about the spillover of momentum effects from stocks to bonds. In the years after, we were able to turn this research into a publication in the Journal of Banking & Finance, and this also forms the basis of Chapter5. Although my name is on this dissertation, I could not have written it without all the support and feedback others have given me over the years. vi
First of all, I would like to thank my promotor Joost Driessen. We have had many lively discussions, either via Skype or in person, and they all had in common that they usually took much longer than the planned hour. My personal goal with the PhD track was to get better at putting results in an academic context, and I have definitely learned a lot while working on the various papers together. I couldn’t have wished a better promotor, and I look forward to get the articles published. I would also like to thank Frank de Jong, my co-promotor, and the committee members Lieve Baele, Esther Eiling, Frans de Roon and Ton Vorst, for taking the time to review my work and providing me with feedback. You gave me the biggest compliment possible by saying you had enjoyed reading my work. To me, that is at least as important as the concise feedback on the contents. Second, I like to thank all colleagues and interns at Robeco with whom I have worked with over the years. Not only because of all the time we shared at the office, but also for all the fun activities like running, cycling, after-work drinks, table soccer or the combination of the latter two. I want to thank a few in particular as there are too many to mention all. First, David for giving me the opportunity to combine my job with this PhD track and for regularly challenging my work. I want to thank Martin and Johan for all their fixed income expertise offered to me in the first years of my career and to thank Winfried, who seems to be an expert at everything. Whenever I needed a reference, Winfried would dig up exactly the (typically 1970/1980’s) paper I needed. A big thank you to Weili and Wilma for their valuable advice when I needed it. Last but not least, a big thank you to Patrick for being my mentor throughout my career at Robeco. We both share the passion for corporate bonds and quant and I hope we can make the whole world as enthusiastic about this combination as we are. Our paper on corporate bond factor investing, Chapter6 of this dissertation, is one attempt in achieving that goal. Finally, I like to thank my friends and family. In particular, I want to thank my brother Jurgen for his help with the layout of this dissertation, and my parents for their support throughout my studies and career. Most im- portantly, I want to thank my girlfriend Janneke for her continuous support. It has not always been easy to combine the PhD track with other activities. When people asked me how I combined a full-time job with a PhD track, I usually replied, half-joking, “I get one day a week from Robeco and one day a week from Janneke”. This dissertation marks the end of that period.
Jeroen van Zundert Rotterdam, November 5, 2017 Contents
1 Introduction1
2 Beta: The Good, the Bad and the Ugly5 2.1 Introduction ...... 5 2.2 Theoretical framework ...... 8 2.3 Main empirical results ...... 13 2.4 Robustness checks ...... 29 2.5 Conclusions ...... 33
2.A Literature overview VAR models and decompositions ...... 35
Tables and figures ...... 36
3 Are Stock and Corporate Bond Markets Integrated? Evidence from Expected Returns 49 3.1 Introduction ...... 49 3.2 Data ...... 53 3.3 Methodology ...... 54 3.4 Benchmark results ...... 59 3.5 Robustness ...... 65 3.6 Conclusions ...... 70
3.A Data ...... 72 3.B Estimating the probability of default ...... 75 3.C Estimating volatility over long horizons ...... 76 3.D Equity-bond elasticity ...... 77
Tables and figures ...... 79 viii
4 Volatility-Adjusted Momentum 97 4.1 Introduction ...... 97 4.2 Methodology ...... 100 4.3 Empirical results ...... 104 4.4 Corporate bond results ...... 111 4.5 Conclusions ...... 113
4.A Derivation optimal portfolio ...... 115
Tables and figures ...... 119
5 Momentum Spillover from Stocks to Corporate Bonds 127 5.1 Introduction ...... 127 5.2 Literature review ...... 128 5.3 Data ...... 130 5.4 Results ...... 132 5.5 Robustness checks ...... 141 5.6 Conclusions ...... 145
5.A Distance-to-Default computation ...... 146
Tables and figures ...... 147
6 Factor Investing in the Corporate Bond Market 159 6.1 Introduction ...... 159 6.2 Data and Methodology ...... 161 6.3 Defining factors in the corporate bond market ...... 164 6.4 The benefits of allocating to factors ...... 167 6.5 Strategic allocation to factors in a multi-asset context . . . . . 171 6.6 Conclusions and implications ...... 173
Tables ...... 175
Bibliography 181 Chapter 1
Introduction
Global capital markets play a key role in today’s world economy by connect- ing savers who do not need their wealth right away to those who can utilize the wealth productively, such as companies (for example to build factories, or acquire another company) and governments (to improve public infrastruc- ture, invest in health care, etc.). Savers supply capital via their savings and investments accounts, but also via their pension savings. In the academic world, a large debate is ongoing since the 1960’s on whether capital markets are efficient (Malkiel and Fama, 1970). If markets are efficient, prices of assets should reflect all available information. In this dissertation, I study the efficiency of stock and corporate bond markets, i.e. the two primary means of financing for companies. There are five chapters, each taking a different angle at whether stock and/or corporate bond markets are efficient. First, in Chapter2 individual and aggregate stock returns are decomposed in three components, namely cash flow news, interest rate news and risk pre- mium news. Previous studies, most notably the “good beta, bad beta” paper by Campbell and Vuolteenaho (2004), do not include an interest rate com- ponent. I find that unexpected interest rate changes account for more than 1/3rd of all variation in stock market returns. Using various stock portfolio sorts, I find that exposure to interest rate news, called interest rate beta or “ugly” beta, has a higher price of risk than both nominal cash flow beta and risk premium beta. Second, in Chapter3 I study the link between the prices of stocks and corporate bonds of the same firm. In theory, as both securities are claims on the same firm assets, the prices should be integrated, a notion formalized by Merton (1974). I compare corporate bond implied stock returns, inferred from credit spreads and expected losses due to default, to realized stock returns to determine the level of integration. Surprisingly, there is a strong negative 2 correlation between expected and realized stock returns over the period 1994- 2015: stocks of which the corporate bond implies high returns realize low returns and vice versa. This effect is stronger for firms with higher default risk, as measured by probability of default, leverage or credit rating, and cannot be explained by differences in the pricing of risk factors in stock and bond markets, limits to arbitrage or liquidity premiums. Third, in Chapter4 I study a refined version of the standard cross-sectional momentum strategy, called volatility-adjusted momentum. In particular, while the standard methodology to create a momentum portfolio, originating from the work of Jegadeesh and Titman (1993), ignores differences in volatility amongst the individual assets, a rational investor would adjust for differences in volatility. Volatility-adjusted momentum differs from standard momentum in three ways: (1) assets are sorted on return-to-volatility, not on raw returns, (2) assets are weighted inverse to their volatility within the portfolio and (3) the portfolios target a constant volatility through time. Empirically, I find that volatility-adjusted momentum has much higher Sharpe ratios (from 0.34 to 1.14) and alphas than standard momentum in the CRSP U.S. stock sample spanning the 1927-2015 period. Moreover, for corporate bonds I find a sim- ilar Sharpe ratio of 1.04 when applying volatility-adjusted momentum, while standard momentum does not even reveal a significant momentum premium in corporate bonds. Fourth, in Chapter5 I focus on the effect of momentum spilling over from one asset class, equities (stocks), to another asset class, corporate bonds. As in Chapter3, the pricing of stocks and bonds of the same firm should be re- lated, and thus also momentum effects might be related. Like previous studies (Gebhardt, Hvidkjaer, and Swaminathan, 2005), I find that past winners in the equity market are future winners in the corporate bond market. However, I also find that a momentum spillover strategy exhibits strong structural and time-varying default risk exposures that cause a drag on the profitability of the strategy and lead to large drawdowns if the market cycle turns from a bear to a bull market. By ranking companies on their firm-specific equity return, instead of their total equity return, the default risk exposures halve, the Sharpe ratio doubles and the drawdowns are substantially reduced. Finally, in Chapter6 I study the presence of the size, value, momentum and low-risk anomalies, well-known in the equity literature, in corporate bond markets. I find empirical evidence that these anomaly portfolios generate economically meaningful and statistically significant alphas in the corporate bond market. As the correlations between the single-factor portfolios are low, a combined multi-factor portfolio benefits from diversification between the factors: it has a lower tracking error and a higher information ratio than the individual factors. The results are robust to transaction costs, alternative CHAPTER 1. INTRODUCTION 3 factor definitions, alternative portfolio construction settings and the evaluation on a subsample of liquid bonds. Finally, allocating to corporate bond factors has added value beyond allocating to equity factors in a multi-asset context. In summary, each of the chapters covers both stock and corporate bond markets, either by explicitly studying the relationship between the two (Chap- ters3 and5), studying the same anomaly in both markets (Chapter4) or by porting concepts well-known in one market, i.e. interest rate risk in bond markets and factors in equity markets, to the other market (Chapters2 and 6). 4 Chapter 2
Beta: The Good, the Bad and the Ugly
2.1 Introduction
The relation between stock prices and interest rates is a key theme in finance, which has been approached from various angles by existing work. Some studies focus on understanding the correlation between (aggregate) stock returns and interest rates (for example Baele, Bekaert, and Inghelbrecht, 2010 and Baker and Wurgler, 2012). Other studies investigate whether interest rates contain predictive information for stock returns (Fama and Bliss, 1987; Cochrane and Piazzesi, 2005). Yet another stream of articles investigates whether interest rate factors are priced in the cross-section of stock returns (for example Fama and French, 1993). Finally, recent work has introduced the concept of equity duration and how this affects the cross-section of stock returns (Dechow, Sloan, and Soliman, 2004; Weber, 2016). In this paper we aim to deepen our understanding of the relation between stock prices and interest rates by using the present value approach of Campbell and Shiller (1988). This approach splits stock returns into cash flow news and discount rate news, and many researchers have applied this framework to understand stock returns. The starting point of our analysis is to decompose the total discount rate into the (nominal) risk-free interest rate and the equity risk premium. Aggregate stock returns are then decomposed into cash flow news, risk premium news, and interest rate news. Our main focus is to extend the “good beta, bad beta” approach of Camp- bell and Vuolteenaho (2004), who posit a two-factor model with stocks having a beta for cash flow news and a beta for discount rate news instead of a single (CAPM) beta. By decomposing discount rate news into risk-free rate 6 news and risk premium news, we obtain a three-factor model. Using various portfolio sorts, we show that splitting up the CAPM beta into a cash flow beta, risk premium beta and interest rate beta generates an improved fit of the cross-section of stock returns. In particular, we find that the interest rate beta carries the highest price of risk, which is why we refer to this beta as the ugly beta. Our approach is as follows. Following existing work (for example Campbell and Vuolteenaho, 2004; Campbell, Polk, and Vuolteenaho, 2010), we use vec- tor autoregressive (VAR) models for stock returns, interest rates and predict- ing variables (price-to-earnings ratio, term spread, small stock value spread, and default spread for the aggregate stock market; book-to-market and return- on-equity for the stock-specific model) to identify these components. We es- timate these VAR models both for market returns and for individual stock returns, which allows us to make statements about both the time series and cross-section of stock returns. We focus on the U.S. stock market for the pe- riod 1927-2015. We use a quarterly frequency for these VAR models, and thus use the 3-month T-bill rate as the risk-free rate. This framework generates a range of new insights on the relation between stock returns and interest rates. As mentioned above, our main analysis is that we decompose the CAPM beta into a cash flow beta, risk premium beta and interest rate beta and study the pricing of these betas. Campbell and Vuolteenaho (2004) call the cash flow beta a “bad” beta and the risk pre- mium beta a “good” beta, because an Intertemporal CAPM predicts that the transitory risk premium variation should carry a lower price of risk. We call the interest rate beta an “ugly” beta because it turns out to have a price of risk that is even higher than the price of cash flow risk, while the price of risk premium risk is small and insignificant. We obtain this result using a standard Fama-MacBeth approach, applied to portfolio sorts on value, size, volatility, and various risk exposures. We can statistically reject the two-factor “good beta, bad beta” model in favor of our three-factor model with a sep- arate “ugly” interest rate beta. While the “good beta, bad beta” model can explain up to 26.5% of the variation in portfolio returns, the addition of the “ugly” beta raises this to 31.8%. Our analysis also delivers other insights. First of all, when decomposing the stock market return into cash flow, equity risk premium and interest rate news, we find that interest rate news accounts for more than 1/3rd of the total variation in market returns. The impact of interest rate movements is large because of the persistence of interest rates. We also estimate the relevance of interest rate shocks at the individual stock level, and find it is small: most of the individual stock price movements are due to cash flows news, in line with Campbell, Polk, and Vuolteenaho (2010). CHAPTER 2. BETA 7
Second, we find that interest rates and risk premiums exhibit a negative relation. In fact, an orthogonal positive shock to interest rates drives down the total discount rate: while a 1% increase in the interest rate has a direct dis- counting effect of about -7% on stock prices, this is more than compensated by an indirect effect of interest rates on risk premiums, which corresponds to a stock return effect of 9% when interest rates increase by 1%. This in- direct effect occurs because interest rates negatively predict risk premiums according to our estimates. Hence, the popular idea that the total discount rate varies one-to-one with the interest rate is not supported by our results. This also suggests that interest rate exposure cannot properly be estimated by accounting-based measures such as equity duration (Dechow, Sloan, and Soliman, 2004; Weber, 2016), as those measures directly measure the duration of the cash flows of the firm, ignoring indirect risk premium effects. Third, we find that nominal interest rate news comoves somewhat nega- tively with nominal cash flow news and with equity risk premium news. Given that movements in nominal interest rates are to a substantial degree driven by changes in expected inflation (Brennan and Xia, 2002), this suggests that stocks might provide a hedge for expected inflation risk, though the size of these hedging effects is limited. This limited inflation hedging capability is in line with most existing work on the inflation hedging aspects of stock invest- ments (see for example Bekaert and Wang, 2010). Our paper relates to various existing streams in the literature. First, it builds on the stream of literature using the present-value framework of Camp- bell and Shiller (1988). Campbell (1991) finds that monthly U.S. unexpected aggregate stock market returns are driven by real cash flow news and discount rate news in similar proportions, while Campbell and Ammer (1993) find that discount rate news is the dominant driver. Campbell and Vuolteenaho (2004) find that exposure to the cash flow news, “bad beta”, is significantly higher priced than exposure to discount rate news, called “good beta”. Using this beta decomposition, they are able to explain the size and value anomalies. Vuolteenaho (2002) is the first paper to decompose individual stock returns, and finds that cash flow news is the main driver of firm-level stock returns. Campbell, Polk, and Vuolteenaho (2010) combine the market and firm-level decompositions to examine the sources of the good and bad betas. They find the firm-level cash flows to be the main driver of the different exposures of value and growth stocks to aggregate discount rate and cash flow news. Second, our paper is linked to papers estimating interest rate sensitivity of stocks more directly. Baker and Wurgler (2012) add the excess return of long- term government bonds to the CAPM model to find that “bond-like stocks”, which are stocks of large, mature, low-volatile, profitable and dividend paying firms, comove more strongly with government bonds than other stocks. Maio 8 and Santa-Clara (2017) employ a similar model to explain the dispersion in average returns of CAPM anomalies (value, return reversal, equity duration, asset growth, inventory growth), and find short-term interest rates relevant for pricing cross-sectional equity risk premia. Weber (2016) uses balance sheet data to construct a measure of duration of the firms cash flows, similar to a Macaulay duration for bonds. He finds that the stocks with a high duration earn lower returns than short-duration stocks in the cross-section. The remainder of this paper is organized as follows: Section 2.2 describes the three-way stock return decomposition employed as well as the VAR model to estimate the individual return components. Section 2.3 discusses the main results, while Section 2.4 contains robustness checks. Section 2.5 concludes.
2.2 Theoretical framework
2.2.1 Decomposing stock returns into shocks The methodology of this paper builds on the log-linearization of asset returns introduced by Campbell and Shiller (1988) and Campbell (1991). Let rt+1 denote the log return of a stock from time t to time t + 1. In general, this can be written as: Pt+1 + Dt+1 rt+1 = log (2.1) Pt where Pt and Pt+1 are the prices of the stock at time t and t + 1 respectively, and Dt+1 any dividend paid out to the investor at time t + 1. Denote with lowercase letters log variables, i.e. log Dt = dt and log Pt = pt. Taking a first-order Taylor expansion around the mean of the log dividend-price ratio d − p leads to the following equation
rt+1 ≈ ρ (pt+1 − dt+1) + ∆dt+1 − (pt − dt) (2.2) where ρ ≡ 1 is the linearization constant. Iterating forward, taking 1+ed−p T expectations and ruling out rational bubbles, i.e. limT →∞ ρ pt+T → 0, results in ∞ κ X p ≈ + ρj ((1 − ρ)d − r ) (2.3) t 1 − ρ Et t+j+1 t+j+1 j=0
1 where κ ≡ − log(ρ) − (1 − ρ) log ρ − 1 and Et the investors expectation at time t. Equation 2.3 shows that the price of an asset is high when future cash flows, i.e. dividends, are high and future returns are low. CHAPTER 2. BETA 9
Substituting Equation 2.3 into Equation 2.2 leads to a two-way decompo- sition of unexpected asset returns into cash flow (CF ) news and discount rate (DR) news:1
∞ ∞ X j X j rt+1 − Et [rt+1] = (Et+1 − Et) ρ ∆dt+1+j − (Et+1 − Et) ρ rt+1+j j=0 j=1
Nt+1 = NCF,t+1 − NDR,t+1 (2.4)
This equation shows that unexpected high stock returns can be generated in two ways: either future cash flows are expected to be higher at t + 1 than the expectation at time t, or expectations of future discount rates are lowered. Thus, if the expectations on future cash flows do not change, any gain today must be offset by losses in the future and vice versa. In the literature various decompositions of stock returns are used. Camp- bell and Ammer (1993) and Campbell and Mei (1993), for instance, split stock returns into real dividends, real interest rates and excess stock returns. Campbell and Vuolteenaho (2004), Campbell, Polk, and Vuolteenaho (2010) and Campbell, Giglio, and Polk (2013) split stock returns into real dividends and excess stock returns, while assuming real interest rates remain constant. In Appendix 2.A we provide a detailed literature overview of the various return decomposition models, including details on the estimation inputs. Before we turn to the specification used in this paper, we first discuss a more general decomposition of stock returns which captures the various two-way and three-way specifications in use. We rewrite Equation 2.4 into a four-way decomposition by decomposing the log discount rate rt+1 into the 2 expected log rate of inflation πt, a log inflation premium θ , the ex-ante real expected log real-interest rate yt and the log excess stock return et+1:
nom real rt+1 = yt + et+1 = πt + θ + yt + et+1 (2.5)
nom where yt is the log of the nominal short-term risk-free rate. By definition the return of the nominal risk-free asset is known at the moment of investing. As a result, the risk-free component in the total discount rate rt+1 which spans the period from t to t + 1, is given by the risk-free rate observed at time t.
1From here on, we assume the approximate equalities in Equations 2.2 and 2.3 hold exactly. 2Brennan and Xia (2002) show that the nominal risk-free rate may also contain a risk premium for unexpected inflation shocks, which we denote by θ. We assume this term is constant over time from here on, which it also is in the Brennan-Xia model. Thus this does not affect the decomposition of unexpected returns. 10
nom nom Hence we denote the nominal risk-free rate as yt , not as yt+1 . Equation 2.4 becomes:
Nt+1 = rt+1 − Et [rt+1] = et+1 − Et [et+1] ∞ ∞ X j nom X j = (Et+1 − Et) ρ ∆dt+1+j − (Et+1 − Et) ρ πt+j j=0 j=1 (2.6) ∞ ∞ X j real X j − (Et+1 − Et) ρ yt+j − (Et+1 − Et) ρ et+1+j j=1 j=1
= NCF nom,t+1 − NΠ,t+1 − Nyreal,t+1 − Ne,t+1
nom nom where the second equality follows from Et [yt ] = yt , and the third equality from substituting Equation 2.5 into Equation 2.4. Note that dt+1+j has been nom written as dt+1+j to emphasize the difference with real dividends. If it is assumed that real interest rates remain constant as is commonly assumed (Campbell and Vuolteenaho, 2004; Campbell, Polk, and Vuolteenaho, 2010; Campbell, Giglio, and Polk, 2013), i.e. Nyreal,t+1 = 0, then the shock not due to excess returns Ne,t+1 represents a shock to real dividends NCF real,t+1 = NCF nom,t+1 − NΠ,t+1. Equation 2.6 is of theoretical interest, but difficult to implement. This is because real rates and expected inflation are not directly observable.3 In this study we therefore focus on measuring the total nominal interest rate component of stock returns, and focus on the term NCF nom,t+1 = NCF real,t+1 + NΠ,t+1. The decomposition that we use in our empirical analysis is given by
Nt+1 = rt+1 − Et [rt+1] = et+1 − Et [et+1] ∞ ∞ X j nom X j nom = (Et+1 − Et) ρ ∆dt+1+j − (Et+1 − Et) ρ yt+j j=0 j=1 (2.7) ∞ X j − (Et+1 − Et) ρ et+1+j j=1
= NCF nom,t+1 − Nynom,t+1 − Ne,t+1
From Equation 2.7 it follows that we obtain three components: interest rate risk Nynom,t+1, equity premium risk Ne,t+1 and nominal cash flow risk NCF nom,t+1.
3Only for our small part of our sample period inflation-indexed bonds and inflation swap data are available. CHAPTER 2. BETA 11
Thus our results for cash flow cannot be directly compared to those of Camp- bell and Vuolteenaho (2004), Campbell, Polk, and Vuolteenaho (2010) and Campbell, Giglio, and Polk (2013), since in these articles it is assumed that real interest rates are constant in order to determine a real cash flow compo- nent. In contrast, we allow interest rates to vary and decompose stock returns into nominal cash flows, nominal interest rates, and risk premiums. In Section 2.3.4, we do, however, use Equation 2.6 to interpret our results for nominal interest rates and to relate our results to existing work.
2.2.2 VAR methodology
To implement the three-way decomposition in Equation 2.7, we follow Camp- bell (1991) by using a vector autoregressive (VAR) model. In this method, first P∞ j nom P∞ j the terms Et [et+1], (Et+1 − Et) j=1 ρ yt+j and (Et+1 − Et) j=1 ρ et+1+j are estimated. Next, given the realization et+1, we can compute the total shock Nt+1 and back out the nominal cash flow shock NCF nom,t+1 as a resid- ual by Equation 2.7. This process has the advantage of only having to estimate expected stock excess returns and nominal interest rates, not the dynamics of the dividend process. Engsted, Pedersen, and Tanggaard (2012) show that if the VAR model is properly specified, it makes no difference whether cash flow news is computed directly or backed out as a residual. In the robustness checks, we estimate a VAR model with less state variables and find that the re- sults are very similar; they are thus not driven by the inclusion of a particular state variable. We assume the data originates from a first-order VAR model given by
Xt+1 = AXt + ut+1 (2.8) where Xt+1 is a m×1 vector of state variables with et+1 as the first element and nom yt+1 as the second element, and m − 2 remaining variables which are included to predict the first two components. A, also called the companion matrix, is a m × m matrix of the parameters of the model to be estimated and ut+1 an m × 1 vector containing the VAR innovations corresponding to Xt+1. We do not include a constant but instead we demean the state variables prior to estimation. The VAR model is used to decompose the stock market return into cash flow, interest rate, and risk premium components. As discussed below, we also estimate a VAR model at the firm level to obtain these components for individual stocks. From the process in Equation 2.8, the total, cash flow, interest rate and 12 equity risk shocks can be computed as follows:
0 Nt+1 = e1 ut+1 0 −1 Ne,t+1 = e1 ρA (I − ρA) ut+1 0 −1 (2.9) Nynom,t+1 = e2 ρ (I − ρA) ut+1 0 0 −1 0 −1 NCF nom,t+1 = e1 + e1 ρA (I − ρA) + e2 ρ (I − ρA) ut+1 where e1 is an m × 1 vector with the first element set to 1 and the remaining elements to zero, e2 an m × 1 vector with the second element set to 1 and the remaining elements to zero, I the m × m identity matrix and A the estimated companion matrix. The term (I − ρA)−1 captures the persistence of a shock in a particular state variable. Variables which have a small direct impact but are very persistent can thus have high impact as stocks are long-term assets (i.e., ρ is typically close to one as dividends are usually 10% or less). Note that there is a difference between the excess return shock term Ne,t+1 and the interest rate shock term Nynom,t+1: the former has an additional multiplication with A due to the risk-free rate being known at the moment of investing.
2.2.3 Beta decomposition As in Campbell and Vuolteenaho (2004), we define the market beta of stock i as the covariance of total firm-specific shocks with contemporaneous total market shocks, scaled with the variance of total market shocks:
Covt (ei,t+1 − E [ei,t+1] , eM,t+1 − E [eM,t+1]) βi,M = Vart (eM,t+1 − E [eM,t+1]) (2.10) Cov (N ,N ) = t i,t+1 M,t+1 Vart (NM,t+1) where ei,t+1 is the log excess return of stock i and eM,t+1 is the log excess return of the market. Following our decomposition of firm-specific and market- wide total shocks in equity risk premium news −Ne, interest rate news −Ny, and cash flow news NCF nom , we can split this single beta into three betas by splitting the market total shock into excess return, interest rate and nominal cash flow shocks:4 βi,M = βi,eM + βi,yM + βi,CF M (2.11) where the subscripts denote the firm-specific (suffix i) and market (suffix M) shock components used to compute the covariance. Note that the signs of
4One could also split the firm-specific shocks into three components using the firm VAR model, yielding a total of 9 betas. In this study we focus on the betas involving total firm-specific shock to a particular market shock. CHAPTER 2. BETA 13 the risk premium and interest rate news terms have been flipped following Equation 2.7. A positive innovation of the shock means that current stock prices rise. The interest rate news component thus depends negatively on interest rate shocks, and can be interpreted in a similar way as a bond return.
2.3 Main empirical results
We first present results for an aggregate VAR model, to decompose market returns into risk premium, interest rate and cash flow news. We then turn to a firm-level VAR model to estimate the stock-specific shocks, which allows us to study the cross-section of the news components and study the various betas defined in Equation 2.11. In particular, we study whether these betas carry different prices of risk. Our dataset draws from several databases. First, for the stock data, we use data from the Center of Research in Security Prices (CRSP) monthly stock files. We limit ourselves to common equity (share codes 10 and 11) and to stocks traded on the NYSE, AMEX or NASDAQ. This dataset contains stock prices, stock returns (including dividends) and shares outstanding. Second, we merge the Compustat Annual database, which contains accounting data for most publicly traded U.S. stocks, into our dataset. As the Compustat database does not contain book values of equity prior to 1952, we use hand-collected book equity values as used in Davis, Fama and French (2000). This data is provided on the website of Kenneth French.5 Third, we source the CRSP US Treasury and Inflation files for the 3-month, 1-year and 10-year (nominal) US Treasury yields. The 10-year yield is provided from 1941 onwards. To cover the 1927-1940 period, we prepend this series with the Long-Term U.S. Government Securities series provided by the Federal Reserve Bank of St. Louis.6 The sample data is on a quarterly frequency and runs from February 1927 to October 2015, spanning 355 quarters.
2.3.1 Aggregate VAR We estimate our aggregate VAR on a quarterly frequency, following Campbell, Giglio, and Polk (2013) and Campbell et al. (2017). The quarterly frequency of the data is a compromise of statistical strength and ability to do more granular analyses (i.e., rolling model estimation; sub period analyses) on the one hand, for which a high data frequency is needed, and the focus on longer term relationships between variables on the other hand, as stocks are typically
5http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data library.html 6https://fred.stlouisfed.org/series/LTGOVTBD 14 thought of as long-term assets, making the estimation of month-to-month dynamics less relevant. We use six state variables in our VAR model, covering the models used in earlier studies since Campbell and Vuolteenaho (2004). See Appendix 2.A for details.
First, the excess log return of the market eM,t is the difference between annual log return on the CRSP value-weighted stock index and the annual log risk-free rate. When computing the return of the individual stocks, delisting returns are taken into account when available to prevent survivorship bias. nom Second, the log risk-free rate yt is the log yield of the 3-month US Treasury bond in fractions. We pick the 3-month rate as the data is on a quarterly frequency, and transform it from an annual rate to a quarterly rate in order to match the investment horizon.
Third, we compute the term yield spread TYt as the difference between the ten-year fixed maturity rate on US Treasuries and the 3-month rate. The term yield spread is quoted in percentages. This variable is included because the term yield spread is known to predict long-term bond excess returns (Fama and Bliss, 1987). Keim and Stambaugh (1986) and Campbell (1987) point out that stocks are also long-term assets, hence TY might also forecast stock excess returns. Moreover, the yield curve tracks the business cycle, and expected stock market returns are likely to vary along the business cycle.
Fourth, we include the log smoothed Shiller price-to-earnings ratio PEt as the ratio of the current stock price to the trailing 10-year earnings of the S&P500 index (Campbell and Shiller, 1988). This ratio captures fluctuations in market valuations, with high (low) ratios indicating the stock market to be expensive (cheap), and thus implying lower (higher) long-run returns in the future. We source this ratio from the website of Robert Shiller.7
Fifth, we include the small-stock value spread VSt. To compute the value spread, we use the 2x3 size and book-to-market portfolios provided by Kenneth French. These portfolios are constructed by rebalancing the portfolios at the end of June of year t by taking the intersection of two size groups and three book-to-market groups. The size breakpoint is the median NYSE size; for book-to-market, the 30% and 70% percentiles of NYSE book-to-market values are used, where the book and market values are from December of year t − 1. Within the small cap stocks, we take the difference of the logs of the book-to- market ratio of the high and the low book-to-market portfolio as our measure.
Sixth, we include the default spread DEFt, computed as the difference between the log yield on Moody’s BAA and AAA bonds. The series are ob-
7http://www.econ.yale.edu/∼shiller/data/ie data.xls CHAPTER 2. BETA 15 tained from the Federal Reserve Bank of St. Louis.8 This variable is included because the default spread reflects both aggregate default probabilities, which should be related to future cash flows, as well as a credit risk premium, which should be related to the equity market risk premium. Table 2.1, panel A, presents the coefficient estimates of the aggregate VAR model.9 The first row shows that the predictability of quarterly market excess returns is limited, as the adjusted R-squared is only 6.6%. However, this is still higher than other studies on quarterly frequency, with Campbell, Giglio, and Polk (2013, Table 3) obtaining a R-squared of 4.0% and Campbell et al. (2017, Table 2) obtaining a R-squared of 3.4% in slightly shorter data samples. As noted by Cochrane (2007), even such a low R-squared might generate substantial variation in risk premiums, because the predicting variables are persistent. The low R-squares do indicate that most deviations of the long-run mean are unexpected. Still, we find statistically significant predictive power for most variables. Past excess returns have a positive and significant (t-statistic of 2.98) impact similar to Campbell, Giglio, and Polk (2013). Nominal interest rates have a negative sign (t-statistic of -2.43), which conflicts with the popular idea that expected total stock returns can be decomposed in the risk-free rate and a constant risk premium (Sharpe, 1964). The negative coefficient suggests that interest rate changes are partly mitigated by risk premium changes in the opposite direction. For the other three variables, we find that the term- yield spread has a positive sign as expected but is statistically insignificant. The value spread has a significant negative (t-statistic of -1.82) impact on future stock returns, while the default spread coefficient is close to zero and statistically insignificant. These results are similar to Campbell, Giglio, and Polk (2013). In the second row, the dynamics of nominal interest rates are given. We find that last quarter’s interest rate is by far the dominant driver with a coefficient of 0.9866 (t-statistic of 46.40). This shows that interest rates are persistent and only slowly mean-revert. Lagged values of the excess market return also have a positive and significant impact on the risk-free rate, while other variables have limited statistical power. In the remaining rows, the dynamics of the term yield spread, the price-to-earnings ratio, the value spread and the default spread are given. We find that the own-lags have a large and highly significant impact; the cross-variable terms have a more limited impact. In panel B, the variance-covariance matrix of the news terms is given on the left. These are derived by computing the shock vector ut+1 each period fol-
8https://fred.stlouisfed.org/categories/32348 9We find the modulus. i.e. the maximum absolute eigenvalue, of the VAR to be 0.9834, below 1. Thus the VAR is stationary. 16 lowing from the VAR model coefficients, and subsequently applying Equation 2.9 to split the excess return shocks into equity risk premium, interest rate and cash flow news. We set ρ to 0.95 at an annual level for this computation. The results are insensitive to modest variations in ρ. As mentioned above, we flip the signs of the risk premium and interest rate news terms for ease of interpretation: an increase in −Ne,t+1, −Nynom,t+1 or NCF nom,t+1 raises the current stock price, and a positive covariance means that the impact on the current stock price is in the same direction. We find that the largest compo- nent is risk premium news with a variance of 0.0060 out of a total of 0.0066, thus representing 90.8% of the total variation. Campbell, Giglio, and Polk (2013) find a similar figure.10 Of the remaining variation interest rate news and nominal cash flow news are approximately equally important, with 36% and 40% respectively. Together, these contributions add up to 167%, which can be attributed to the negative covariance terms. Interestingly, we find modest evidence for stocks being real assets. If we assume that real interest rates are constant, shocks in the nominal interest rate are driven by shocks in the expected inflation. If stocks are real assets, then an increase in inflation expectations should be offset by an increase in nominal cash flow expectations, by a decrease of future excess returns or a mix of the two. We find that the two covariance terms with interest rate news are indeed negative (Table 2.1, Panel B), but these offsetting effects are not very large. The correlations, listed on the right, equal -0.18 and -0.38 for risk pre- mium news and cash flows news, respectively. These results can be compared to the literature on the limited inflation hedging capacity of stock market in- vestments. For example, Bekaert and Wang (2010) document low correlations between stock market returns and inflation rates for many countries. Bekaert and Engstrom (2010) show that increases in expected inflation coincide with increases in equity risk premiums, while our results suggest a negative relation. The correlation between news on future nominal cash flows and future excess returns is small with -0.15. This suggests that, on a quarterly frequency, stock returns are driven almost independently by updated future excess returns and changes in future cash flow expectations. One should be careful interpreting the finding that interest rate news drives as much of the variation in stock returns as nominal cash flow news, as the variance-covariance matrix does not show what the initial shock, i.e. “trigger”, is for shocks in stock returns, only where these triggers accumulate: changes in expected future excess returns, future interest rates or future cash flows. The finding that interest rate “news” is a large component of stock returns should
10They report only the correlations and standard deviations. From these, it can be inferred that the ratio of variances of risk premium news to total news is 88.9%. CHAPTER 2. BETA 17 thus not be interpreted as stocks being very sensitive to contemporaneous interest rate changes. Indeed, the VAR coefficients suggest that future excess returns partly offset the future interest rate changes. Instead, it shows that updated expectations on future interest rates, irrespective of the “trigger”, have contributed substantially to the total variation in stock returns. Panel C of Table 2.1 provides more insight in which shocks are driving the news terms. On the left, the correlations between the residuals of the VAR state variables and the news terms are given. Unsurprisingly, shocks in the excess return correlate positively with risk premium news, interest rate news as well as cash flow news. A shock in the interest rate correlates negatively with risk premium news, implying that a positive interest rate shock causes future excess returns to be lower, raising current stock prices. As interest rates are persistent, future interest rates will also be higher, leading to lower current stock prices and hence we observe a strong negative correlation of -0.76 between interest rate shocks and future interest rate news. The correlation with cash flow news is relatively small, with a value of just -0.06. Shocks in the term yield spread TY correlate mainly with interest rate news: an increase in the term yield spread can be caused by higher 10-year rates, lower 3-month rates or a positive combined effect of the two. If the 3-month rate is lower, it means lower future interest rates, leading to positive interest rate news. For the price-to-earnings ratio, most of the shock comes from changing prices, as the 10-year earnings only slowly updates. The correlation with the excess return residual is 0.85 (not reported). Thus, it mainly correlates with risk premium news. The correlations for the value spread are relatively small, while an unexpected increase in the default spread, which typically happens at the start of a recession, leads to higher future risk premia, lower expected interest rates and lower future expected cash flows, in line with economic intuition. The right part of panel C shows the analytical mapping functions as defined in Equation 2.9. These functions should be interpreted as marginal effects: given a one-unit shock in the respective state variable, and no shocks in any of the other state variables, it shows the distribution over the three news terms. Thus by construction, the terms either add up to 1 for excess return (as the three news terms add up to the excess return shock), or to 0 for all other state variables (as we force the excess return shock to be zero). Still, these functions provide insight in where shocks in particular state variables accumulate. For excess returns, we find that the majority of an orthogonal positive shock would be taken as positive news on future cash flows, with a relatively small positive effect on future excess returns, and a small negative effect on future interest rates. For interest rate shocks, we find a strong negative effect of -29.14 (t- statistic of -1.41): a higher interest rate means future interest rates will also be 18 higher, lowering current stock prices. This is a direct discounting effect, and one can thus interpret the coefficient of -29.14, or -29.14/4 = -7.3 annually, as a duration effect. This direct discounting effect is, however, more than offset by lowered future excess returns (coefficient of 36.59), due to the fact that the interest rate negatively predicts risk premiums in the VAR model. It should be noted that the coefficients for interest rate shocks are not all statistically significant though. For the term yield spread, value spread and default spread the absolute t-statistics do not exceed 1.20, meaning they do capture one of the news terms specifically. For the PE ratio, we find that an increase leads to a strongly negative cash flow. This is an intuitive result, as we impose that the excess return residual has to be zero, and thus by definition the price will change very little. The PE ratio can thus only increase due to lowered earnings. The lower the earnings of companies, the lower the expected future dividends will be. In Figure 2.1, the news terms are plotted through time. For the sake of readability, the series have been smoothed by taking a exponentially weighted trailing average. In line with the covariances in the table, we observe little co-movement between the three news terms. The causes for up and down- turns of the stock market are thus diverse. During the Great Depression (end 1920s), the cash flow shock was extremely negative, contributing substantially to this crisis. However, also equity risk was discounted more heavily. The only other period where both cash flow and risk premium news were very negative together was during the Great Financial Crisis (2008). Campbell, Giglio, and Polk (2013) find similar patterns for this “hard time”. In the other “hard time” they document, namely the early 2000s, it was mainly due to investors increasing future equity premia leading to current stock prices dropping, af- ter a long period of compressing equity risk premia in the second half of the 90s. For interest rate news, the largest losses occurred in the early 80s. This is mainly driven by strong upward shocks in interest rates at the time. We conclude that there is sufficient variation in the news terms through time to separately estimate the price of each.
2.3.2 Firm-level VAR Besides the aggregate VAR, we also estimate a firm-level VAR to be able to derive shocks on the firm and portfolio level. These shocks can then be used to estimate betas of portfolios using Equation 2.11. To estimate the firm-level VAR, we include the same state variables as for the market-level VAR, except that we add the firm-specific log excess stock 11 return ei,t, log book-to-market ratio bmi,t and log return-on-equity roei,t. 11We do not remove the aggregate versions of the state variables, as we want these to drive CHAPTER 2. BETA 19
The choice for these variables follows Vuolteenaho (2002) and Campbell, Polk, and Vuolteenaho (2010). The log book-to-market ratio is included to capture cross-sectional differ- ences in valuations between stocks, where high (low) ratios indicate higher (lower) future long-run excess returns (Graham and Dodd, 1934). We com- pute the ratio by applying shrinkage prior to taking the log following Camp- bell, Polk, and Vuolteenaho (2010).12 This is necessary, as taking the log of values (close to) zero leads to extreme observations. Therefore, we shrink the book-to-market ratio to 1, with a weight of 10% on the prior and 90% on the observation, resulting in 0.9BEi,t + 0.1MEi,t bmi,t = log (2.12) MEi,t where BEi,t (MEi,t) is the book (market) value of equity of firm i at quarter t respectively. We assume the book value at the close of December of a particular year is available from April the next year onwards. The market value of equity is always the most recently observed value, ensuring the book-to-market ratio changes from quarter to quarter. The log return-on-equity ratio is included to capture the evidence that firms with higher profitability, controlling for their book-to-market ratio, earn higher average stock returns (Haugen and Baker, 1996). We construct the measure as in Vuolteenaho (2002). First, to compute return on equity, we divide last year’s US GAAP earnings to the beginning of last years book value of equity. The earnings and book value of equity are sourced from Compustat. When earnings are missing, the clean surplus formula is computed using the hand-collected book value of equity data from Kenneth French. See Vuolteenaho (2002) for details on the computation. We ensure that potential losses are not larger than the beginning-of-period book value of equity by winsorization of the earnings. Otherwise, the return-on-equity might be below -100%. The log return-on-equity ratio is then computed as: NIi,t−4:t roei,t = log 1 + 0.9 + 0.1yt (2.13) BEi,t−4 where NIi,t−4:t is the net income over the last four quarters, BEi,t−4 the beginning-of-period book value of equity and yt the 3-month T-bill rate. We the interest rate process in the same way as in the aggregate VAR model. Coefficients of aggregate variables (time t + 1) on firm-specific variables (time t) are restricted to zero. 12For the firm-specific returns, we winsorize negative returns at -99.9% to prevent taking logs of zero. 20 shrink the return-on-equity to the risk-free rate to prevent the log transfor- mation from returning extreme values in case the non-transformed return-on- equity is close to -100%. As most firms do not exist throughout the sample, we assume that the companion matrix A is the same for all firms and time periods. In Section 2.4, we relax this assumption and find that our results are not altered in any significant way. We estimate the firm-level VAR model with pooled-panel OLS regressions. To ensure our results are not biased towards the end of the sample due to the strong growth in number of stocks over time, we weight each stock- quarter observation with the inverse of the number of stocks in that particular quarter. This adjustment also ensures that the coefficients for the aggregate variables are independent from the number of observations per time period, and are therefore the same as in the aggregate VAR. Table 2.2 reports the results.13 Panel A reports the dynamics of the firm- specific variables; the rows with the aggregate variables have been omitted for space reasons, as these are the same as in Table 2.1. We find that firm-specific returns are harder to predict than aggregate returns, as the R-squared is only 3.1%, versus 6.1% for the aggregate return. This is not surprising given that individual stock returns exhibit substantial idiosyncratic risk. Of the variables predicting the firm-specific returns well are the log book-to-market ratio, albeit with a t-statistic of only 1.50, and log return-on-equity with a t-statistic of 5.87, confirming the profitability effect. As with aggregate stock returns, past quarter stock market returns have significant positive effect, which is in line with the standard momentum effect of Jegadeesh and Titman (1993), while interest rates and the price-to-earnings ratio have a significant negative impact. Interest rate shocks now have a more negative effect than in the aggregate VAR model. The book-to-market ratio and return-on-equity load strongly on their own lags with coefficients of 0.97 and 0.89 respectively. The variance-covariance matrix of the news terms is listed in Panel B of Table 2.2. The variance of interest rate news is, as it is an aggregate variable, similar to that in the aggregate model.14 However, as the variance of returns on the firm-level is much larger, this represents just 3.9% of the total variance of firm-specific stock returns. The cash flow news variance is 91.0% of the total variance, and the equity risk premium news variance 32.0%. Thus nominal cash flow news is on the firm-level relatively much more important than it is on market level, where it accounted for just 40.2% of the total variation. This
13As with the aggregate VAR model, we find the maximum modulus of the eigenvalues to be 0.9834. 14We equal weight all firm-quarter observations, rather than equal weighting the cross- sections. Hence the variance deviates slightly from the aggregate VAR model, as it empha- sizes recent quarters due to the growth in number of stocks over time. CHAPTER 2. BETA 21
finding has also been documented by Vuolteenaho (2002). The correlations, listed on the right of Panel B, are modest, ranging from -0.35 to +0.05. Panel C reports the impact of shocks in VAR variables on the news terms. The table on the left shows the empirical correlations between the residuals of the VAR model and the news terms. We find similar effects as in the aggregate model for the firm-specific variables. For the market state variables, we find the correlation with (firm-specific) cash flow news to be approximately zero.
2.3.3 Betas of anomaly portfolios
Since its introduction in the 1960s, the Capital Asset Pricing Model (CAPM) has been challenged by numerous “anomalies”, notably the size (Banz, 1981), value (Basu, 1977) and low volatility (Haugen and Heins, 1972) effects. One way to deal with these findings has been to simply add these anomalies as new factors to the model, arguing these anomalies are compensations for some unknown risks to investors. For instance, Fama and French (1992) developed a three-factor model containing the market, size and value factors. However, this approach does not provide a deeper understanding on why these anomalies exist. The beta decomposition derived in Equation 2.11 can provide this deeper understanding. This study is not the first to employ beta decompositions to study CAPM anomalies. Based on an Intertemporal CAPM, Campbell and Vuolteenaho (2004) argue that market cash flow news should carry a higher risk premium than market discount rate news. This is because discount rate shocks are transitory: low returns due to an increase in discount rates today are partially compensated by higher future expected returns. They find that value and small cap portfolios have outperformed growth and large cap portfolios as they have a relatively high exposure to market cash flow shocks (“bad beta”), relative to exposure to market discount rate shocks (“good beta”). Campbell, Polk, and Vuolteenaho (2010) extend this work by documenting that for value portfolios this is mainly driven by a high sensitivity of the portfolio’s cash flows to the market shocks, not the portfolio’s discount rate shocks. Campbell et al. (2017) extend the two-way decomposition to include a premium for volatility. They find that growth not only better hedges declines in future discount rates, but also increases in volatility, hence demanding a lower risk premium. However, none of these studies have analyzed potential differences in interest rate beta as an explanation for differences in risk premiums. To determine the pricing of the three betas we use portfolios as test assets. As in Campbell and Vuolteenaho (2004), our primary set of test assets are the 5x5 size-by-value portfolios. We construct these portfolios by sorting at 22 the end of July15 of each year all stocks in five market cap groups and inde- pendently in five book-to-market groups where the quintiles are based on the NYSE stocks only. Subsequently, we form 25 market-value weighted portfolios based on the intersection of these groups. For the book-to-market ratio, we take the book and market values as of December of the year before. For the three quarters following the rebalance, we maintain the positions unless there are delistings; in that case, the proceeds are reinvested proportionally in the remaining positions. Besides these 25 portfolios, we add five market value weighted portfo- lios based on past 12-month stock return volatility. These are rebalanced each quarter, and are included to explicitly address the question whether low volatility stocks indeed have a higher interest rate exposure. Finally, we add a third set of portfolios. This set is specifically designed to address the concern of Daniel and Titman (1997) that using only portfolios sorted on characteristics known to influence average returns like value and size might lead an asset pricing model to fit the high variation in mean returns to only small deviations in the betas, as the betas might be close to each other. To ensure there is sufficient spread in the betas, we construct 40 risk-sorted portfolios in the spirit of Campbell and Vuolteenaho (2004). First, per stock and per quarter, we estimate the following OLS regression over the past 20 quarters:
nom ri,t = β0+βM rM,t + βy∆y + βTY ∆TYt t (2.14) + βVS∆VSt + βDEF ∆DEFt + i,t where ri,t is the log excess return of the stock over the quarter, the betas are the coefficients to be estimated, and the independent variables are changes in the aggregate state variables used in the VAR model. We exclude the Shiller price- to-earnings ratio PEt from this regression as quarter-on-quarter changes in the PE ratio are almost entirely driven by stock returns, leading to a very high correlation with rM,t. The delta operator indicates the change from the start of the quarter to the end of the quarter, i.e. the same period as over which the stock return is measured. Stocks are first sorted in five groups based on their estimated βM , and then within each of the five groups sorted in two groups for each of the other state variable coefficients estimated in the regression. This yields a total of 40 risk-sorted portfolios, which are market-value weighted and rebalanced quarterly. The total number of portfolios is thus 70. The portfolio
15These portfolios are similar to the ones provided by Kenneth French, but are rebalanced at the end of July instead of June, as we do not have a quarter ending at the end of June. We need to rebalance at the end of a quarter to be able to determine the news terms of the portfolios, hence we rebalance one month later. CHAPTER 2. BETA 23 shocks are obtained by taking the market-value weighted average firm-specific shocks generated by the firm VAR model. As we require 20 quarters of VAR shocks to compute the risk-sorted port- folios, the first quarter we can compute a return and shock runs from May to July 1932. For comparability, we use for all portfolios the betas from May 1932 to October 2015. As Campbell and Vuolteenaho (2004) find substantial differences in betas in the pre- and post-1963 periods (labeled “early” and “modern” sample), we also report results separately for these periods, where the early sample runs from May 1932 to January 1963 (123 quarters) and the modern sample from February 1963 to October 2015 (211 quarters). Table 2.3, panel A, reports the betas for the size x value portfolios. In the full sample we find that small caps tend to have higher risk premium betas, higher interest rate betas but lower cash flow betas than large caps. For the early sample, a similar pattern emerges, but in the modern sample the relation between size and the interest rate beta is flat while the cash flow betas are actually higher. For value stocks, we find that they have lower interest rate betas then growth stocks in the early sample, but that this relation is flat in the modern sample. The risk premium beta results are mixed in the early sample, while in the modern sample value stocks have clearly lower risk premium betas than growth stocks. For the nominal cash flow betas, the results are very consistent: value stocks have higher cash flow betas than growth stocks. Campbell, Giglio, and Polk (2013) find a similar pattern for the real cash flow betas of value stocks in their model. In general, we find clear differences in betas between the early and modern sample for the size x value portfolios, consistent with prior studies. Panel B reports the results for the five volatility portfolios. Not surpris- ingly, due to the close relationship between volatility and beta, the risk pre- mium and cash flow betas tend to increase from low volatility to high volatility, which is consistent across the samples. The exception is the interest rate beta. Measured over the full sample as well as in the two sub samples, low volatility stocks tend to have a higher interest rate beta than high volatility stocks, con- sistent with other studies (Baker and Wurgler, 2012; Maio and Santa-Clara, 2017). However, we also note that this is mainly due to high volatility stocks having a lower interest rate beta, Q1 to Q4 are relatively close to each other. Panel C reports the full sample results for the risk-sorted portfolios. Of particular interest are the portfolios double-sorted on market beta and the interest rate change, as the coefficient on the interest rate change corresponds closely with the interest rate beta. Within the high market beta portfolios, we find that stocks which had, in the past, a high sensitivity to interest rate changes (i.e., high βY in Equation 2.14), indeed also a higher interest rate beta going forward. However, it seems that this effect only exists in market beta 24
Q1 and Q2 portfolios, not in the lower market beta portfolios.
2.3.4 Pricing of the betas In this section, we analyse the cross-sectional pricing of the betas using the 70 portfolios constructed in the previous section. Campbell (1993) derives a discrete time version of the Merton (1973) intertemporal capital asset pricing model (ICAPM) model to show that the expected return of an asset is a linear function of the betas under certain assumption. This pricing model has been employed by for instance Campbell and Vuolteenaho (2004) and Campbell et al. (2017) to estimate the premium on each of the betas. To understand how nominal interest rate risk should be priced in an In- tertemporal CAPM, we consider two extreme cases. First, if real interest rates are constant, all variation in nominal interest rates is due to changes in expected inflation. Then, nominal interest rate news and nominal cash flow news sum up to real cash flow news. Assuming that the ICAPM investor cares about real returns, the ICAPM of Campbell and Vuolteenaho (2004) implies a single price of risk for real cash flow news, and in this case nominal cash flow risk and nominal interest rate risk should thus carry the same price of risk, while risk premium news has a lower price of risk. Campbell and Vuolteenaho (2004) assume that real rates are constant, hence this case corresponds to the way they interpret their results. The alternative extreme case is that inflation is constant. In this case, all variation in nominal interest rates is due to changes in real rates. These real rates directly enter the total discount rate, and hence in this case the price of interest rate risk should equal the price of risk premium risk. In reality, both real rates and inflation vary over time and the pricing of nominal interest rate risk will differ from both cash flow risk and risk premium risk. This is why we mainly focus on a model where all three components have separate risk prices. As discussed below, we do however also include a specification that follows Campbell and Vuolteenaho (2004). In addition, it is important to note that the Intertemporal CAPM is just one justification for why the different components carry different prices of risk. In particular, there are several reasons why investors might care about interest rate risk beyond its effect on stock prices. First, interest rates may affect risk premiums in bond markets, and investors who invest both in bonds and stocks will care about this. Second, pension funds and insurance companies have liabilities that depend on interest rates. Third, interest rate risk may be related to systemic liquidity risk which could in turn affect financial markets beyond stock markets. To estimate the prices of risk, we employ a Fama and MacBeth (1973) CHAPTER 2. BETA 25 procedure using the full-sample equity risk premium, interest rate and cash flow betas by estimating per year the following equation using Ordinary Least Squares: Ri,t = λe,tβi,eM + λyβi,yM + λCF βi,CF M + i,t (2.15) where Ri,t is the simple excess return of portfolio i over the 3-month T-bill rate, and βi,eM (βi,yM , βi,CF M ) the full-sample estimated risk premium (in- terest rate, cash flow) beta of portfolio i. The subscript t denotes the quarter; the lambda coefficients, which represent the return per unit of beta, are aver- aged over time. Standard errors are computed in two ways: first, we report heteroskedasticity and autocorrelation corrected standard errors (Newey and West, 1987).16 Second, we report bootstrapped standard errors in square brackets where we resample quarters with replacement.17 We compute the price of the betas under three assumptions: 1. CAPM all betas are equally priced. This means we are pricing the total (CAPM) beta.
2. GBBB the interest rate and cash flows betas are equally priced (λy = λCF ), the risk premium beta might differ. This is the same setup as in the “Good Beta, Bad Beta” study (Campbell and Vuolteenaho, 2004).
3. unrestricted all three betas priced separately. For each of the three assumptions, we also re-estimate Equation 2.15 including a constant. There are two interpretations to the results with constant versus those without the constant. First, including a constant can be viewed as a model misspecification test. The betas should capture all risks, and thus no significant positive or negative excess return, captured by the constant, should remain. Alternatively, it can also be viewed as a lighter test on the model, as we no longer force the model to price both the equity premium as well as the cross-sectional differences between the 70 portfolios, but only the cross-section. A priori, we would expect a small and probably insignificant premium for the risk premium beta, as it represents transitory risk, while for the sum of the interest rate and nominal cash flow betas (i.e. GBBB model) we would expect a highly positive premium if inflation risk is an important component
16We ignore the uncertainty in the estimation of the betas and news terms themselves. A common way to incorporate uncertainty in the beta estimates is to use the Shanken (1992) correction. However, we do not derive the betas from a multivariate regression as is common (Cochrane, 2001), but instead compute covariances directly as implied by the Campbell- Shiller return decomposition. When we use multivariate regressions to estimate the betas, we find similar values, and the Shanken correction factor amounts to 1.84. 17We thus maintain the cross-sectional correlation structure between the portfolios and the factors. 26 of interest rate risk, as it adds up to the real cash flow beta representing permanent risk (Campbell and Vuolteenaho, 2004). For the interest rate beta alone we would also expect a positive risk premium, as interest rate changes pose a risk to bond investors. Government bonds are assumed to carry a positive risk premium to compensate for this risk (Fama and Bliss, 1987), and thus one would also expect this premium to appear in stock returns. As discussed above, whether the nominal cash flow beta has a premium similar, lower or higher than the interest rate beta is undetermined. Table 2.4, panel A, reports the full sample results. We first focus on the results without constant. Under the CAPM assumption, we find a positive and significant price of risk of 2.14% per quarter per unit of beta. The R- squared of 11.4%18 indicates that the CAPM beta is not really able to price the portfolio returns. The GBBB model is much better at explaining the returns, increasing the R2 to 26.5%, while the mean absolute pricing error of the portfolios drops from 0.45% to 0.30%. To test whether the constraint λe = λy+CF that the CAPM implicitly imposes versus the GBBB model can be rejected, we employ an F -test. The F -test statistic of 40.54 indicates that the GBBB model is also statistically significantly better at explaining the returns than the CAPM model. It achieves this by setting a low and insignificant price on the risk premium beta and a high premium on the real cash flow beta (βi,yM +βi,CF M ) of 6.63% per quarter. Campbell and Vuolteenaho (2004) have found similar differences in the pricing of the two betas. The unrestricted model reveals, however, significant differences in the pricing of the interest rate beta versus the nominal cash flow beta. Although both are positive and significant, the interest rate beta has a much higher price of risk than the nominal cash flow beta. The F -statistic of 12.63 indicates that the unrestricted model is a statistically significant improvement over the GBBB model. Also, the R2 increases, and the mean absolute error declines slightly. We also find the unrestricted model to be significantly better than the CAPM model (F - statistic of 30.05). As mentioned above, there are several reasons why interest rate risk may carry a higher price of risk than cash flow risk. The three right-most columns of Table 2.4, panel A, show the same anal- yses but then with a constant included. Clearly, the CAPM model without
18The adjusted R-squared is computed as follows per quarter:
n − 1 P 2 R2 = 1 − i i P ¯ 2 n − k i(Ri − R) where n is the number of portfolios, k the number of regressors, i the residuals, Ri the return of portfolio i and R¯ the average over the portfolios returns. This allows negative R-squares as we force the constant to be zero. Subsequently, the quarterly R-squares are averaged over time. CHAPTER 2. BETA 27 constant only loads on the CAPM beta to explain the equity risk premium; the CAPM beta does not explain cross-sectional dispersion in returns, in line with literature on the relation between beta and return to be relatively flat (Haugen and Heins, 1972). Also with a constant, we find the GBBB and unrestricted models to be significant improvements over the CAPM model (F -statistics of 3.58 and 6.24 respectively), although the gain in R2 is less strong than without constant. For the GBBB model, we find that part of the loading on the real cash flow beta is shifted to the constant, which is positive and just significant (t-statistic of 1.70), implying misspecification of the model without constant. For the unrestricted model, the loadings on the three betas hardly change when the constant is included; the constant itself is close to zero and insignificant. We can thus not reject the hypothesis that the model without constant is correctly specified. Moreover, also with a constant included we find the unrestricted model to be statistically stronger than the GBBB model in pricing the portfolios (F -statistic of 6.24). Panel B reports the results for the pre-1963 period. In general, we find that the GBBB and unrestricted models substantially improve upon the CAPM model when no constant is included. However, once a constant is included, we find hardly any statistical evidence of improved pricing of the GBBB and unrestricted models over the CAPM model. Campbell and Vuolteenaho (2004) come to a similar result, and point out that their real cash flow betas are during this period approximately a fixed proportion of the total beta across the test assets, making it hard for the asset pricing test to assign prices. We find that in our early sample this proportion ranges from 5% to 26%, whereas in the full sample it ranges from 15% to 43%. Thus there indeed seems to be less variation relative to the full sample, but variation in the relative betas certainly exists. Panel C reports the results for the post-1963 period. For the CAPM model, the results are very similar: the constant subsumes the total beta coefficient. Both the GBBB and unrestricted models prove significant improvements over the CAPM model (F -statistics of 6.83 and higher), although the improvement of the unrestricted model over the GBBB model is now smaller and not always significant. The premium on the interest rate beta is a much smaller 3.24% (1.09%) versus the full sample estimate of 10.32% (10.07%) and the early sample estimate of 15.04% (4.09%) for the model without (with) constant. Still, in the model with constant the interest beta is statistically significant and positive with a t-statistic of 1.79. Figure 2.2 plots the prices of risk through time for the final specification. For the sake of readability, the prices are smoothed using an exponentially weighted average with half-time of 12 quarters. It is immediately clear from the figure that the prices vary through time, but are also relatively smooth. While the risk premium (excess return) 28 beta is usually close to zero or slightly negative, both cash flow and especially interest rate premiums vary strongly and are mostly positive. For the interest rate beta, 1999-2000, i.e. the dot-com bubble, is an exceptional period with a very negative price of risk. The interest beta premium tends to comove with the cash flow premium (correlation of 33%), and to a lesser extent with the risk premium (correla- tion of 15%). Still, there are also periods where the correlation breaks down between interest rate and cash flow prices, such as the early 1950s and late 1960s. So far, the assessment of the various asset pricing models has focused on statistical evidence: are the betas priced differently? Another way to compare the models is to check whether a model is better able to price the anomaly portfolios. As a measure, we compute per portfolio the pricing error, that is, the difference between the quarterly realized return and the fitted return of the pricing model, and subsequently take the average over the absolute values of errors. The smaller this mean absolute error (MAE), the better the pricing model is able to explain the returns of the anomaly portfolios. Table 2.4 reports in the bottom two rows of each panel the MAE. The first row (“MAE1”) is computed over all 70 portfolios, i.e. the exact same set as on which the prices are calibrated; the second row (“MAE2”) reports the MAE over the 25 size x value portfolios plus the 5 volatility portfolios. Full sample, the improvement over all 70 portfolios is small for the three- beta model over the GBBB model. However, for the anomaly portfolios we find clear evidence the pricing improves when the interest beta is included, especially in the cross-section: the MAE drops from 0.4620% to 0.4178%. It also shows that still a relatively large portion of the errors remains. Over the early sample, the MAE is close to unchanged, and for the modern sample we observe a modest decrease in the pricing error. In Table 2.5 the pricing errors for selected anomaly portfolios19 are dis- played per model specification. Although we observe that in general the errors indeed tend to shrink to zero when the interest beta is added, this is typically by a small magnitude. This holds for the size, value as well as the low volatility effect. To conclude, there is clear statistical evidence that the interest rate beta is priced differently from the cash flow and risk premium beta, but we find the economic significance to be limited when it comes to pricing anomaly portfolios.
19Of the 25 size x value portfolios, only the Q1/Q3/Q5 combinations are shown to save space. The other portfolio results are available upon request. CHAPTER 2. BETA 29
2.4 Robustness checks
2.4.1 Reduced VAR model
In Section 2.3 we based all results on the aggregate VAR model with six state variables. Although the selection of state variables follows the literature well (see Appendix 2.A for a discussion), the inclusion of particular variables might drive our findings. Therefore we redo the analyses with a reduced VAR model. For the absolute minimum number of variables, we need to have the market excess return and the risk-free rate to estimate the two respective news components, as well as a variable which includes the current stock price, oth- erwise the decomposition no longer holds (Engsted, Pedersen, and Tanggaard, 2012). Amongst the six state variables, the Shiller PE ratio fulfills this role. The Shiller PE also happens to be the strongest of the four remaining state variables in predicting future market returns in the full VAR model (Table 2.1). Hence we re-estimate the VAR model with the market excess return, the risk-free rate and the price-to-earnings ratio. Panel A of Table 2.6 contains the results of the estimation. We find that the adjusted R2 for the market return decreases slightly versus the full model, from 6.57% to 5.88%. For the risk-free rate and the price-to-earnings ratio it is virtually the same. The estimated transition matrix coefficients are also very similar. This indicates that there are no major changes compared to the full model. In panel B, the covariance matrix of the news terms is shown. Compared to the full model, the covariances between the news terms change only modestly. Of the variance of the three terms, only the nominal cash flow changes from 0.0026 to 0.0018. In panel C, the correlations between the full model and reduced model news terms are listed. The correlation between the excess return components is very high at 0.97. The correlations for the other two news terms amount to 0.75. They are thus highly correlated, but not perfect. To assess whether this has a meaningful impact on the pricing of the betas, the analysis in Table 2.4 is repeated with the reduced VAR model news terms in panel A of Table 2.7. We find that the results are similar or even stronger: the interest beta has a very high and significant premium, and is statistically different from the nominal cash flow premium. Interestingly, we also find that the inclusion of the interest rate beta im- proves the pricing performance compared to the GBBB model: not only is the statistical evidence stronger (higher F -statistics), but also the mean absolute error decreases by about 25%, whether a constant is included (specification 6 vs. 5), or not (specification 3 vs. 2). This is large compared to the reductions found for the full model. 30
2.4.2 Out-of-sample estimation of the news terms Underlying our main results is the assumption that the VAR model properly captures investors’ expectations. In the previous robustness check, we checked the sensitivity of the results for the inclusion of particular state variables, but this still assumes that the relationships between the state variables are constant over the sample and known in advance to investors. To incorporate a dynamic relationship between the state variables, stock returns and interest rates in particular, as well as to prevent foresight we re- estimate the VAR model in a rolling fashion. At each quarter, we use the previous 40 quarters, i.e. 10 years, to estimate the transition matrix and apply this to the next quarter to compute the news terms. If we would do this without any constraints, the transition matrix becomes non-stationary during some periods, as indicated by a modulus being greater than 1. To prevent this situation, we apply the following restrictions:
• We use the reduced VAR model to prevent too many coefficients to be estimated over relatively few data points.20
• We explicitly restrain the interest rate dynamics by requiring the same long-term average as in the full model as well as the same own-lag coef- ficient of 0.9834. If we would not do this, the own-lag coefficient can be greater than 1 at times, as interest rates are only stationary in the very long run, not per se on a 10-year horizon.
• To prevent non-stationarity, we explicitly restrain the modulus of the transition matrix to be 0.99 or smaller.
To incorporate the restrictions, we estimate the rolling VAR with the General- ized Method of Moments (GMM). Like Campbell, Giglio, and Polk (2013), we use Hansen, Heaton, and Yaron (1996) continuously updated (CUE) GMM as it has finite-sample advantages over standard GMM. For comparability with the full sample estimations, we include the first 10 years (February 1927 to January 1937) by using the coefficients estimated over this period. Each quarter, we re-estimate the VAR coefficients and apply these out- of-sample to compute the news terms. We find that the total shock variance increases from 0.0067 for the full sample estimated VAR model to 0.0073 for the rolling VAR. The increase is relatively modest, given that the fit is no longer in-sample. If we analyze the decomposition in more detail, we find
20The full VAR model has 42 coefficients, as it has 6 state variables and per state variable 7 coefficients (a constant and 6 coefficients on the lagged state variables). The reduced VAR model has only 12 coefficients, of which we effectively fixate 2 (see second point), leaving 10 free coefficients to be estimated. CHAPTER 2. BETA 31 that the risk premium variance decreases from 0.0060 to 0.0044, while the other two variances increase. This suggests that the lack of foresight makes the changes in expected excess returns more muted, attributing more of the total shock variance to interest rates (from 0.0023 to 0.0037) and nominal cash flows (from 0.0026 to 0.0060) instead. Comparing the shock terms head-to- head, the total shocks have a correlation of 94%, while the individual news terms correlations range from 75% to 84%. This suggests that the in-sample estimation procedure does not lead to overfitting. Panel B of Table 2.7 reports the asset pricing results for the full sample. In contrast to the full-sample VAR model results, we observe that nominal cash flow premium is no longer statistically larger than zero in the unrestricted model, while the price attached to the interest rate beta becomes even larger. Therefore, the evidence of the added value of splitting the real cash flow beta into an interest rate beta and a nominal cash flow beta is even stronger, with the F -statistics being 85.50 and 12.82 for the models without and with a constant respectively. Also economically, the gain of including the interest rate beta is large: the mean absolute pricing error decreases dramatically, with the strongest decrease for the anomaly portfolio under the specification without constant: from 0.5748% to 0.3017%. Thus by allowing one additional free parameter, the pricing error is almost halved.
2.4.3 Characteristic-dependent firm VAR model In the main results we have assumed that the companion matrix A is the same for all firms. As we are employing characteristics-sorted portfolios in our asset pricing test, we are making the strong assumption that the dynamics of the individual stocks are the same across these portfolios. To alleviate this concern, we modify the firm VAR model by adding interaction dummies based on the characteristics. Specifically, each period we divide the universe in three equal sized groups based on the book-to-market ratio (market capitalization or 1bmQ1 1bmQ1 12-month historic volatility) to create the dummies i,t and i,t+1 , where the superscript indicates the characteristic on which the sort is based (i.e., bm for book-to-market, me for market capitalization equity and vol for volatility) and whether the stock is on this characteristic in the top (Q1) or bottom (Q3) tertile, where the ordering is such that high values of the characteristic are assigned to Q1 and low values to Q3. We take the middle group Q2 as the base group and hence do not include a dummy for this group. To ensure the groups are equal sized, we compute the dummies for t and t + 1 every quarter again. Thus entrants and leavers in the universe do not bias the sizes of the groups. To prevent the number of parameters to become very large, we only interact nom the dummies with the interest rate state variable yt , and interaction terms 32
nom are only driven by the interest rate variable yt and the interaction terms based on the same characteristic. Table 2.8 reports the results. For space reasons, the rows and columns with the aggregate state variables have been removed, with the exception of the interest rate column, which is the only aggregate variable to drive the inter- acted variables. As we have three characteristics and two additional terms per characteristic, there are six new state variables in total. These state variables, being firm-specific, are constrained to only explain firm-specific variables, as are the other firm-specific variables. We first discuss the bottom 6 rows of the matrix, which describe the dynamics of the interaction terms, and then the impact on other state variables. From the coefficients, we observe that the interaction terms are positively nom1bmQ1 nom autocorrelated. For instance, yt+1 i,t+1 has a coefficient of 0.11 on yt , 0.75 nom1bmQ3 on its own lag and -0.11 on yt i,t . Thus if a stock belongs to the top book- nom1bmQ1 to-market tertile Q1, next quarters value for the state variable yt+1 i,t+1 is nom expected to be 0.11 + 0.75 = 0.86yt , which is positive. In words: if a stock is this quarter in the top book-to-market group, it likely is next quarter as well. The coefficient of 0.75 indicates how persistent this likeliness is.21 On the other hand, if the stock currently belongs to the bottom group, next quarters expected value is 0.11-0.11, which is close to zero. Thus, the data tells us that it is unlikely a stock moves from the bottom to the top group from one quarter to the next. The opposite is also true: the coefficient for a stock nom1bmQ3 currently being in the top group ending up in the bottom group yt+1 i,t+1 equals 0.09-0.09, which is zero as well. This persistence is also visible for the market cap groups, which have higher own-lag coefficients than the book- to-market interaction terms. This indicates that the market capitalization dummies are more persistent than the book-to-market dummies. For volatility, we also find persistence, but the own-lag coefficients are the lowest of the three characteristics, indicating that stocks transition groups fastest along the volatility dimension. With these interaction terms, we can assess the impact on the other state variables. For the news decomposition the stock-specific log excess return nom ei,t+1 and the interest rate dynamics yt+1 are of particular importance. As the interest rate is an aggregate variable, it is not affected by the inclusion of firm-specific interaction terms. For the log excess return, we find that the effects are relatively modest and statistically insignificant for the book- to-market, but not for market cap and volatility: t + 1 excess returns of low
21These coefficients also pick up the persistence in interest rates. As the coefficient for interest rates on its own lag is 0.87 is less than one, the coefficients of the interaction terms cannot directly be interpreted as probabilities of switching from one group to another. CHAPTER 2. BETA 33 market cap (volatility) stocks are more (less) negatively exposed to an interest rate shock in year t than middle market cap (volatility) stocks. For volatility, stocks in the Q1 group are also more negatively exposed. This has implications for the news terms: if market participants expect a difference in the reaction of high volatility stocks to interest rate shocks versus low volatility, an interest rate shock will lead to different expected excess re- turns versus the firm-VAR model employed in the previous section. Results of the beta decomposition of the portfolio as well as the pricing of the betas show, however, very small differences (results are available upon request). The rea- son is that although the expected excess returns can now be estimated slightly more precisely, firm-specific excess returns remain very hard to predict.22 For comparison, the standard deviation of interest rate residuals is 0.16%, whereas the standard deviations of the firm-specific excess return residuals in the base case firm VAR model is 24.99%. Thus the difference in coefficients between high and low volatility of 2.18 (0.70 + 1.48) is relatively small in comparison to the difference in the volatilities of the two state variables.
2.5 Conclusions
Since the introduction of the present-value decomposition of asset returns by Campbell and Shiller (1988), a number of studies have exploited this break- down of stock returns into several components to better understand stock return dynamics, both in aggregate as well as in the cross-section. Although there has been active debate on the interest rate exposure of stock returns, and given the important role interest rates play in financial markets in gen- eral, to the best of our knowledge no other studies have separately estimated a nominal interest rate component of stock returns. Instead, most studies opt for estimating real cash flow and total discount rate shocks. In this study, we disentangle interest rate shocks and nominal cash flow shocks, and find that unexpected interest rate shocks, i.e. “interest rate news”, explains as much of the variation in stock returns as nominal cash flow news. Although interest rates have much lower variation than stock excess returns, we find that the high persistence of interest rates leads to sizeable effects. On the firm level, interest rates explain only 3.3% of the total variation, as interest rates, in contrast to firm-specific excess returns and cash flows, are
22By construction, adding explanatory variables in an OLS framework can only reduce the unexplained variation in the excess returns, not increase. The adjusted R-square, which corrects for the number of regressors included, shows a small increase from 3.05% (Table 2.2) to 3.30% (Table 2.8), indicating limited additional explanatory power of the six interaction terms. 34 identical for all firms in a given period, and are thus unable to explain any cross-sectional variation by construction. We also analyze the sensitivity of anomaly portfolios to interest rate news. We find that growth stocks, large stocks and low volatility stocks have higher interest betas than value, small and high volatility stocks, but these relation- ships can vary through time. Allowing the prices of the three betas (equity risk premium beta, interest rate beta and nominal cash flow beta), to vary leads to a considerable improvement in explaining the returns of the anomaly portfolios. The price of interest rate beta is estimated to be positive, and is in most tests significantly different from the nominal cash flow beta price. Thus treating the two betas separately reveals important distinctions in interest rate exposures and pricing previously not observed. CHAPTER 2. BETA 35
2.A Literature overview VAR models and decom- positions
In the academic literature, multiple VAR models are in use. Table 2.9 sum- marizes the main features of those models. The model in this study draws primarily on the model of Campbell, Giglio, and Polk (2013). For the sample period, we are limited by the CRSP and COMPUTSTAT datasets, hence we start in February 1927. The quarterly frequency of the data is a compromise of statistical strength and ability to do more granular analyses (i.e., rolling VAR model estimation and rolling betas) on the one hand for which monthly or quarterly data is needed and the focus on longer term relationships between variables on the other hand, as stocks are typically thought of as long-term assets, making the estimation of month-to- month dynamics less relevant. We include all state variables used in the four most recent papers except stock return variance as we are not interested in a volatility beta. Campbell et al. (2017, Table 1, panel B) find that this variable has very limited and statistically insignificant forecasting power for the other state variables, thus excluding it is unlikely to affect our results. Campbell et al. (2017) is the only study of the four to use real returns; in this study we opt for excess stock returns rather than real returns for reasons explained in the beta decomposition section. For the firm-VAR model, there are less studies to draw on; see Table 2.10. Both Vuolteenaho (2002) and Campbell, Polk, and Vuolteenaho (2010) use annual data with excess stock returns, book-to-market ratio (for the cross- sectional value effect) and a profitability measure. 36 ρ nom nom M,CF M,CF N N nom nom M,y M,y Adj R2 N N − − t M,e M,e N N DEF 0.04570.6678 -0.03140.0303 0.2456 -0.0144 0.1858 -0.9133** -0.2161 0.1821 -0.0194 0.8373*** 0.0773 -0.1812 -0.1454 -0.0036 0.1036 -0.1000 -0.1454 -0.3761 0.0514 -0.1812 0.0484 -0.3761 − − 36.5884 -29.1380 -7.4504 (0.7438) (-0.8652)(1.5347)(0.1052) (-0.2567) (0.9012) (1.1972) (-2.3928) (-0.6781) (1.5050) (-0.2731)(1.2192) (-1.4091) (7.2464) (-0.3496) (-0.0230) (1.3694) (-0.5871) nom nom +1 t t +1 +1 +1 +1 t M,e t t +1 nom t N M,y VS M,t y TY VS PE e − N M,CF DEF Corr/std Functions − N t PE nom nom t TY M,CF M,CF N N nom nom nom t M,y M,y y N N − − -statistics (in brackets) and adjusted R2’s. The state variables are constructed as described in Section t M,e M,e M,t N N e 0.7527 0.1856 0.2711 0.82760.2224 0.3031 0.2489 -0.1948 -0.2359 0.4649 -0.7579 -0.0558 0.0781 -1.8002*** -0.0116*** 0.0250*0.0060 0.8934*** -0.0007 0.0687*** 0.9451 -0.0006 -0.2985 0.4353 -0.0543 -0.4166 0.4136 -0.5170 -0.1534 -2.6343 0.8526*** 0.0489 0.0149-0.0006 0.1739*** 0.8150 -0.0009 0.0026 -0.0007 0.0023 -0.0009 − − 0.0025** 0.9866*** 0.0002 -0.0002 -0.0002 -0.0004* 0.9592 -0.4952** 5.2462** 0.0044 0.0005 0.2168*** 0.8284*** 0.8347 0.1805*** -1.7159** 0.00170.3635*** -0.0385*** -0.0350* -0.3705 0.0008 0.0051 0.9726*** 0.0657 -0.0261 0.0049 0.9563 nom nom +1 +1 t t +1 +1 +1 +1 +1 +1 +1 +1 t t M,e t t t t +1 +1 nom nom t t N M,y M,t M,t y y TY TY VS VS PE PE e e − N M,CF DEF DEF Var/covar − N (value spread) (1.3961) (-2.7611) (-2.6449) (1.8294) (41.2596) (6.1499) (risk-free rate) (2.0108) (46.3992) (1.4542) (-0.7391) (-0.6614) (-1.7818) (default spread) (-2.2153) (2.3751) (0.4068) (0.0122) (2.6595) (14.5067) (term yield spread) (-0.3984) (-0.4715) (25.7095) (0.6449) (0.1380) (2.7688) (excess market return) (2.9840) (-2.4291) (0.3817) (-2.9441) (-1.8299) (0.0485) Corr. shocks and news (price-to-earnings ratio) (5.8013) (-0.5445) (1.0823) (73.2478) (-1.2581) (0.2969) Panel A: VAR coefficients Panel C: Shocks and news Panel B: News Table 2.1:Panel Aggregate A VAR reports model the2.2 OLS estimatedand coefficients, demeaned priordeviations, to diagonal estimation. (right)correlations, Panel of on B the the reports right newsquantities, the the terms. the marginal empirical delta-method effects covariance1946 ) Panel of (Cram´er, matrixof is a C (left) 0.95 used. state and is reports variable Significance assumed the shock the at in correlations, as the mapping the implied off-diagonal 90%, computation by of terms, 95% the of a and VAR and the model. standard shock 99% news To in levels terms. compute are a Sample the indicated standard period state with errors from variable *, of February to ** these 1927 and derived the to *** October news respectively. 2015. terms: A left the empirical CHAPTER 2. BETA 37 ρ Adj R2 t DEF nom nom -statistics of these derived t t VS M,CF M,CF N N nom nom t M,y M,y PE N N − − t M,e M,e N N TY 0.1420 -0.3491 -0.1937 0.3409 -0.01940.1418 -0.3215 1.0596 -0.0314 0.2456 -0.1104 -1.3051** -0.1937 0.0469-0.0003-0.0561 0.2395 1.0003*** 0.0561 -0.0751 0.1036 -0.0285 -0.3491 0.0494 -0.0469 -0.2830 0.1858 0.0972 − − 69.5739 -29.1380 -40.4360 (1.5265) (-0.2884)(1.2496) (-1.5476) (-1.5294)(1.3663) (-0.8969) (-1.0068)(1.3738) (-1.1795) (0.9482) (-2.0142) (-0.0248)(-0.6895)(-5.4423) (75.2215) (0.6895) (5.4423) (-0.2640) (1.0720) (-0.1035) (-0.4797) (1.1042) (0.1643) -0.2784*** 0.2784*** nom nom +1 +1 +1 t +1 +1 +1 +1 +1 t M,e t t +1 i,t i,t nom nom t t i,t N M,y M,t y y e TY VS PE e − N M,CF bm roe DEF Corr/std Functions − N M,t e nom nom i,t roe M,CF M,CF N N -statistics (in brackets; clustered by firm-quarter as suggested by Petersen)) ( 2009 and adjusted R2’s t nom nom i,t M,y M,y bm N N − − M,e M,e i,t e N N 0.0202 -0.0025 -0.0066 0.0103 0.9666*** 0.0076 -0.2694*** 0.7733 -0.00660.3123 0.0272* 0.0442 0.0434* 0.8538 -0.00570.6265 0.9192 0.2091 -0.0329 0.4173 -0.8056 -0.0780 0.68480.0380 0.2630 0.2278 -0.1220 0.0216 -0.3739-0.3795 -0.0411 -0.0078 -0.5607 0.2691 -0.4449 0.3715 -0.0088 -0.0082 0.0055 0.0326*** 0.2968*** -2.9006*** -0.0008-0.0025 -0.0648***-0.0066 0.0024 -0.0406 0.0006 0.0006 0.0023 0.0573 0.0305 -0.0971 0.3899 -0.0264 − − 0.0284*** -0.0104*** 0.8943*** -0.0419*** -0.7415*** -0.0062*** -0.0252*** 0.0144*** -0.0098*** 0.7872 nom nom +1 +1 +1 +1 +1 t +1 +1 +1 +1 +1 +1 t M,e t t +1 i,t i,t i,t i,t nom t i,t i,t N M,y M,t y e e TY VS PE e − N M,CF bm bm roe roe DEF Var/covar − N (excess return) (-0.6280) (1.4995) (5.8730) (3.4916) (-3.2505) (-0.1379) (-4.2080) (-1.4825) (0.1258) (return-on-equity) (7.7799) (-7.5487) (94.5513) (-3.5549) (-5.9849) (-7.9441) (-8.6599) (3.1787) (-3.8935) (book-to-market ratio) (0.8173) (225.9875) (1.6249) (-3.4163) (0.9400) (-1.2154) (1.9002) (1.6827) (-0.3176) Panel A: VAR coefficients Panel B: News Panel C: ShocksCorr. and news shocks and news of a pooled-panel regression.the aggregate For space VAR reasons modeldeviations, only (Table diagonal the2.1 ). results (right) Panelcorrelations, for of B the on the reports firm-specific the news the state right empirical variables terms. the are covariance marginal shown; matrix Panel the effects (left) C other and of reports lines the a are the correlations, state identical off-diagonal mapping variable to terms, shock of those and in as a standard implied shock by in the a VAR state model. variable To to compute the the news terms: left the empirical Table 2.2:Panel Firm A VAR reports model the OLS estimated coefficients, quantities, the delta-method)1946 (Cram´er, of is 0.95 used. is Significance assumed at in the the 90%, computation 95% of and the 99% news levels terms. are Sample indicated period with from *, February ** 1927 and to *** October respectively. 2015. A 38
Table 2.3: Beta decomposition portfolios Portfolios are constructed as described in Section 2.3.3. The table reports per portfolio the beta to the three market news terms generated by the VAR model of Table 2.1. Betas are reported over the full sample period (“Full”) which runs from May 1932 to October 2015, as well as for the early sample (“Early”; May 1932 - January 1963) and the modern sample (“Modern”; February 1963 - October 2015). To save space, we only report Q1, Q3 and Q5 for the 5x5 size x value portfolios; for the risk-sorted portfolio we only report the full sample results. Panel A: size (ME) x value (BE/ME) portfolios ME Low Q3 High BE/ME Low Q3 High Low Q3 High Low Q3 High
Full βi,eM 1.38 1.07 1.17 1.02 0.86 0.96 0.73 0.64 0.86 βi,yM 0.24 0.21 0.16 0.18 0.16 0.17 0.14 0.11 0.12 βi,CF M 0.00 0.09 0.13 0.12 0.17 0.21 0.10 0.17 0.23 Early βi,eM 1.55 1.31 1.47 1.07 1.03 1.24 0.79 0.73 1.23 βi,yM 0.35 0.26 0.20 0.17 0.12 0.13 0.10 0.05 0.02 βi,CF M -0.26 -0.07 0.02 -0.01 0.10 0.14 0.05 0.15 0.24 Modern βi,eM 1.22 0.85 0.90 0.97 0.70 0.71 0.66 0.55 0.53 βi,yM 0.15 0.16 0.13 0.18 0.20 0.21 0.18 0.17 0.20 βi,CF M 0.22 0.24 0.24 0.24 0.23 0.27 0.15 0.20 0.23
Panel B: volatility portfolios Vol Low Q2 Q3 Q4 High
Full βi,eM 0.52 0.78 1.00 1.15 1.37 βi,yM 0.16 0.11 0.14 0.14 0.10 βi,CF M 0.09 0.18 0.20 0.25 0.28 Early βi,eM 0.63 0.86 1.05 1.16 1.28 βi,yM 0.11 0.06 0.10 0.10 0.06 βi,CF M 0.01 0.15 0.13 0.19 0.27 Modern βi,eM 0.42 0.71 0.95 1.14 1.44 βi,yM 0.20 0.17 0.18 0.18 0.14 βi,CF M 0.17 0.21 0.26 0.29 0.28
Panel C: risk-sorted portfolios (full sample only) βM High Q2 Q3 Q4 Low βY High Low High Low High Low High Low High Low
Full βi,eM 1.26 1.11 0.96 0.89 0.82 0.75 0.64 0.55 0.55 0.41 βi,yM 0.18 0.13 0.17 0.14 0.10 0.13 0.14 0.16 0.16 0.14 βi,CF M 0.15 0.24 0.13 0.20 0.17 0.17 0.11 0.14 0.11 0.16
βM High Q2 Q3 Q4 Low βTY High Low High Low High Low High Low High Low
Full βi,eM 1.23 1.13 0.92 0.94 0.80 0.78 0.63 0.58 0.56 0.42 βi,yM 0.18 0.13 0.16 0.14 0.11 0.12 0.15 0.16 0.14 0.15 βi,CF M 0.15 0.25 0.15 0.17 0.16 0.18 0.10 0.16 0.12 0.17
βM High Q2 Q3 Q4 Low βVS High Low High Low High Low High Low High Low
Full βi,eM 1.21 1.16 0.95 0.93 0.80 0.80 0.60 0.60 0.53 0.46 βi,yM 0.13 0.17 0.15 0.17 0.11 0.12 0.14 0.17 0.16 0.14 βi,CF M 0.23 0.19 0.19 0.13 0.19 0.17 0.19 0.09 0.14 0.16
βM High Q2 Q3 Q4 Low βDEF High Low High Low High Low High Low High Low
Full βi,eM 1.16 1.19 0.90 1.00 0.73 0.88 0.53 0.72 0.43 0.60 βi,yM 0.16 0.15 0.16 0.16 0.12 0.11 0.15 0.18 0.17 0.13 βi,CF M 0.16 0.24 0.13 0.19 0.15 0.21 0.11 0.15 0.09 0.22 CHAPTER 2. BETA 39
Table 2.4: Pricing of betas We employ Fama and MacBeth (1973) regressions by regressing each quarter the simple returns of the 70 portfolios (25 size x value, 5 volatility, 40 risk-sorted) in excess of the risk-free rate on the full sample betas to market risk premium shock, market cash flow shock and market interest rate as reported in Table 2.3. In round brackets are the Fama-MacBeth standard errors, corrected for autocorrelation and heteroscedasticity (Newey and West, 1987); in square brackets bootstrapped standard errors. In the first column we restrict all beta coefficients to be equal to each other (“CAPM”), and in the second column we restrict the cash flow and interest rate betas to be equal to each other (“GBBB”) and in the final column we leave all coefficients free (“unrestricted”). The adjusted R2 is the average adjusted R2 of the cross-sectional regressions. The F -tests below the columns show the test-statistic whether the model in that column (i.e. no longer imposing the equality of particular coefficients) is significantly better either the CAPM model (“F -test vs CAPM”) or the “Good Beta, Bad Beta” model (“F -test vs GBBB”). The mean absolute error (MAE) is the mean of the absolute values of the difference of the actual and fitted quarterly mean return of the 70 portfolios (“MAE1”) or only the 30 anomaly portfolios (“MAE2”). Stars denote the significance at the 10% (*), 5% (**) and 1% (***) level. Results are reported for the full sample period which runs from May 1932 to October 2015 (Panel A), as well as for the early sample May 1932 - January 1963 (Panel B) and the modern sample February 1963 - October 2015 (Panel C) separately. no constant constant model CAPM GBBB unrestricted CAPM GBBB unrestricted specification 1 2 3 4 5 6 Panel A: full sample (May 1932 - October 2015) constant 0.0168*** 0.0103* 0.0005 (2.9390) (1.6974) (0.0671) [3.3810] [1.2919] [0.0534] λe (price βi,eM ) 0.0214*** 0.0057 0.0026 0.0077 0.0053 0.0026 (4.6029) (0.7920) (0.3464) (1.1303) (0.7292) (0.3721) [4.7885] [0.5608] [0.2807] [1.1690] [0.6148] [0.3019] λy (price βi,yM ) 0.0214*** 0.0663*** 0.1032*** 0.0077 0.0350** 0.1007*** (4.6029) (4.4299) (4.1507) (1.1303) (2.1209) (2.6487) [4.7885] [3.2998] [3.8297] [1.1690] [1.0768] [1.9732] λCF (price βi,CF M ) 0.0214*** 0.0663*** 0.0491*** 0.0077 0.0350** 0.0479*** (4.6029) (4.4299) (3.2349) (1.1303) (2.1209) (2.9874) [4.7885] [3.2998] [1.7823] [1.1690] [1.0768] [1.5030] Adj. R2 11.4% 26.5% 31.8% 28.3% 32.0% 36.6% F -test vs CAPM 40.54*** 30.05*** 3.58** 4.36** F -test vs GBBB 12.63*** 6.24** MAE1 0.4567% 0.3011% 0.2940% 0.3081% 0.2961% 0.2938% MAE2 0.5515% 0.4343% 0.4161% 0.4941% 0.4620% 0.4178%
Panel B: early sample (May 1932 - January 1963) constant 0.0151* 0.0128* 0.0126* (1.9163) (1.6832) (1.6550) [1.7299] [1.5297] [1.4453] λe (price βi,eM ) 0.0331*** 0.0250** 0.0136 0.0209 0.0189 0.0181 (3.6336) (1.9880) (0.9042) (1.6212) (1.3319) (1.3799) [3.8483] [1.8073] [0.7706] [1.6275] [1.1796] [1.0489] λy (price βi,yM ) 0.0331*** 0.0740** 0.1504** 0.0209 0.0422 0.0490 (3.6336) (2.4857) (2.3089) (1.6212) (1.3029) (0.9655) [3.8483] [2.1152] [1.8799] [1.6275] [1.1003] [0.6990] λCF (price βi,CF M ) 0.0331*** 0.0740** 0.1047*** 0.0209 0.0422 0.0454* (3.6336) (2.4857) (3.1902) (1.6212) (1.3029) (1.9353) [3.8483] [2.1152] [2.4594] [1.6275] [1.1003] [1.1054] Adj. R2 9.9% 18.9% 23.8% 25.3% 30.6% 32.7% F -test vs CAPM 12.29*** 8.50*** 2.82* 1.33 F -test vs GBBB 4.14** 0.02 MAE1 0.4504% 0.4000% 0.3800% 0.3400% 0.3266% 0.3258% MAE2 0.4882% 0.4647% 0.4555% 0.4544% 0.4303% 0.4304% 40
(Table 2.4 continued) no constant constant model CAPM GBBB unrestricted CAPM GBBB unrestricted specification 1 2 3 4 5 6 Panel C: modern sample (February 1963 - October 2015) constant 0.0181** 0.0022 0.0066 (2.3889) (0.2882) (0.8120) [2.9504] [0.2337] [0.6327] λe (price βi,eM ) 0.0143*** -0.0030 -0.0056 -0.0004 -0.0030 -0.0066 (2.8922) (-0.3563) (-0.6265) (-0.0548) (-0.3571) (-0.6967) [2.8684] [-0.3171] [-0.5170] [-0.0591] [-0.3443] [-0.6329] λy (price βi,yM ) 0.0143*** 0.0478*** 0.0324* -0.0004 0.0426*** 0.0109 (2.8922) (3.5171) (1.7925) (-0.0548) (2.6014) (0.5235) [2.8684] [3.5449] [1.3941] [-0.0591] [1.7712] [0.2901] λCF (price βi,CF M ) 0.0143*** 0.0478*** 0.0689*** -0.0004 0.0426*** 0.0612*** (2.8922) (3.5171) (3.3217) (-0.0548) (2.6014) (2.8580) [2.8684] [3.5449] [2.0905] [-0.0591] [1.7712] [1.9030] Adj. R2 12.0% 31.7% 33.7% 32.0% 35.4% 36.9% F -test vs CAPM 94.10*** 49.25*** 13.10*** 6.83*** F -test vs GBBB 2.42 3.79* MAE1 0.4846% 0.2804% 0.2712% 0.3333% 0.2829% 0.2764% MAE2 0.6086% 0.4054% 0.3926% 0.4908% 0.4118% 0.3992% CHAPTER 2. BETA 41
Table 2.5: Pricing errors anomaly portfolios under various asset pricing models Pricing errors full sample under the six specifications in Table 2.4. Errors represent quarterly returns, and are defined as realized return minus fitted return. The sample runs from May 1932 to October 2015.
Panel A: size (ME) x value (BE/ME) portfolios ME Low Q3 High BE/ME Low Q3 High Low Q3 High Low Q3 High 1 -1.45% 0.36% 1.43% -0.31% 0.41% 0.81% -0.09% 0.35% 0.43% 2 -0.36% 0.68% 1.93% -0.06% 0.28% 0.62% -0.05% 0.06% 0.20% 3 -0.82% 0.42% 1.93% -0.20% 0.25% 0.66% -0.18% 0.15% 0.45% 4 -0.93% 0.55% 1.75% -0.18% 0.36% 0.96% -0.45% -0.06% 0.40% 5 -0.60% 0.63% 1.87% -0.11% 0.31% 0.81% -0.29% -0.04% 0.30% 6 -0.82% 0.42% 1.93% -0.20% 0.25% 0.67% -0.19% 0.15% 0.45%
Panel B: volatility portfolios Vol Low Q2 Q3 Q4 High 1 0.39% 0.28% -0.52% -0.67% -1.48% 2 0.07% 0.19% -0.48% -0.62% -1.06% 3 -0.19% 0.33% -0.35% -0.35% -0.54% 4 -0.24% 0.07% -0.38% -0.26% -0.78% 5 -0.15% 0.11% -0.41% -0.39% -0.84% 6 -0.19% 0.32% -0.35% -0.34% -0.54% 42 nom M,CF N nom M,y N − M,e N 0.2646 -0.3822 0.0424 0.0763 -0.4632 0.2646 − -0.4632 0.0502 -0.3822 nom nom M,e M,y N M,CF N − -statistics in brackets. The final column reports Corr/std t − N of 0.95 is assumed in the computation of the news ρ Adj R2 nom nom t PE M,CF M,CF N N nom nom nom t M,y M,y y N N − − M,e M,e M,t N N e 0.00580.0009 -0.0018 -0.0008 0.0009 0.9688 0.0018 0.2529 -0.1160 -0.3281 -0.1235 0.7462 − − -0.0018 0.0025 -0.0008 -0.4706 0.7536 -0.0056 0.1825** -1.1146* -0.0343*** 0.0588 0.0031*** 0.9834***0.3575*** 0.0000 -0.3371 0.9586 0.9709*** 0.9562 nom nom nom nom +1 +1 t M,e M,e +1 nom t M,y M,y N N M,t y e PE M,CF M,CF N N − − Var/covar − − N N reduced/full (risk-free rate) (2.5925) (66.8369) (0.0124) (price-to-earnings ratio) (4.9145) (-0.5747) (82.3894) (log excess market return) (2.3275) (-1.9380) (-3.1410) Panel C: Correlation full VAR model news with reduced VAR model news Panel A: VAR coefficients Panel B: News Table 2.6: ReducedPanel A aggregate reports VAR the model OLSthe estimated adjusted coefficients of R2. thedescriptive aggregate The statistics VAR model state on as variables the wellterms) are implied as and news the constructed the terms. as standardat described deviations On the (diagonal). the in 90%, left, Panel Section 95% theterms. C2.3.1 reports empirical and Theand covariance the 99% sample demeaned matrix; correlations period levels on prior with are runs the the to from indicated right, news estimation. February with the terms 1927 *, correlations of Panel to the (off-diagonal ** B October full and 2015. reports VAR *** model. respectively. Significance A CHAPTER 2. BETA 43
Table 2.7: Pricing of betas under the reduced VAR model & under the rolling reduced VAR model We employ Fama and MacBeth (1973) regressions by regressing each quarter the simple returns of the 70 portfolios (25 size x value, 5 volatility, 40 risk-sorted) in excess of the risk-free rate on the full sample betas to market risk premium shock, market cash flow shock and market interest rate. In round brackets are the Fama-MacBeth standard errors, corrected for autocorrelation and heteroscedasticity (Newey and West, 1987; Newey and West, 1994); in square brackets bootstrapped standard errors. In the first column we restrict all beta coefficients to be equal to each other (“CAPM”), and in the second column we restrict the cash flow and interest rate betas to be equal to each other (“GBBB”) and in the final column we leave all coefficients free (“unrestricted”). The adjusted R2 is the average adjusted R2 of the cross-sectional regressions. The F -tests below the columns show the test-statistic whether the model in that column is significantly better either the CAPM model (“F -test vs CAPM”) or the Good Beta, Bad Beta model (“F -test vs GBBB”). The mean absolute error (MAE) is the mean of the absolute values of the difference of the actual and fitted quarterly mean return of the 70 portfolios (“MAE1”) or only the 30 anomaly portfolios (“MAE2”). Stars denote the significance at the 10% (*), 5% (**) and 1% (***) level. Results are reported for the full sample period which runs from May 1932 to October 2015. Panel A reports the results for the full-sample reduced VAR model (see Table 2.6) and Panel B reports the results for the out-of-sample rolling reduced VAR model. no constant constant model CAPM GBBB unrestricted CAPM GBBB unrestricted specification 1 2 3 4 5 6 Panel A: full sample reduced VAR model constant 0.0172*** 0.0109* -0.0001 (3.0228) (1.9371) (-0.0118) [3.4741] [1.6917] [-0.0111] λe (price βi,eM ) 0.0213*** 0.0005 0.0185* 0.0074 0.0013 0.0186 (4.5997) (0.0609) (1.9351) (1.0944) (0.1665) (1.6112) [4.7825] [0.0396] [1.3974] [1.1365] [0.1208] [1.2268] λy (price βi,yM ) 0.0213*** 0.0829*** 0.1608*** 0.0074 0.0455*** 0.1613*** (4.5997) (4.4450) (4.4300) (1.0944) (2.6612) (3.1742) [4.7825] [2.9585] [4.0963] [1.1365] [1.2934] [3.2159] λCF (price βi,CF M ) 0.0213*** 0.0829*** 0.0342 0.0074 0.0455*** 0.0344* (4.5997) (4.4450) (1.6104) (1.0944) (2.6612) (1.9050) [4.7825] [2.9585] [1.0369] [1.1365] [1.2934] [1.0268] Adj. R2 11.3% 24.3% 30.7% 28.3% 32.0% 38.7% F -test vs CAPM 41.08*** 47.26*** 6.80** 31.87** F -test vs GBBB 58.14*** 12.77** MAE1 0.4638% 0.3103% 0.2484% 0.3109% 0.2908% 0.2484% MAE2 0.5619% 0.4254% 0.3187% 0.4977% 0.4462% 0.3183%
Panel B: rolling reduced VAR model constant 0.0178*** 0.0171*** 0.0078 (3.1620) (2.9394) (1.2120) [3.4741] [1.6917] [-0.0111] λe (price βi,eM ) 0.0246*** 0.0466*** 0.0341*** 0.0079 0.0183 0.0239 (4.5960) (3.2997) (2.5792) (1.0229) (1.0907) (1.3505) [4.7825] [0.0396] [1.3974] [1.1365] [0.1208] [1.2268] λy (price βi,yM ) 0.0246*** -0.0188 0.2146*** 0.0079 -0.0107 0.1668*** (4.5960) (-0.8122) (4.6215) (1.0229) (-0.4640) (3.7358) [4.7825] [2.9585] [4.0963] [1.1365] [1.2934] [3.2159] λCF (price βi,CF M ) 0.0246*** -0.0188 0.0302 0.0079 -0.0107 0.0231 (4.5960) (-0.8122) (1.4654) (1.0229) (-0.4640) (1.1952) [4.7825] [2.9585] [1.0369] [1.1365] [1.2934] [1.0268] Adj. R2 11.2% 15.3% 24.3% 28.0% 32.8% 37.8% F -test vs CAPM 4.96*** 48.85*** 1.52 39.55*** F -test vs GBBB 85.50*** 12.82*** MAE1 0.4749% 0.4504% 0.2728% 0.3167% 0.3154% 0.2592% MAE2 0.5752% 0.5748% 0.3017% 0.5042% 0.5062% 0.3079% 44 Adj R2 3 volQ i,t 1 nom t y 1 volQ i,t 1 nom t y 3 nom bmQ i,t 1 M,CF nom t N y nom bmQq i,t 1 M,y N nom t − y 3 meQ i,t M,e 1 N 0.1451 -0.3353 -0.2204 -0.3353-0.2204 0.0493 0.0435 0.0435 0.2412 − nom t y 1 nom nom meQ i,t M,e 1 N M,y − N M,CF Corr/std of 0.95 is assumed in the computation of the news terms. The sample period runs nom t − N y ρ nom t y (11.3013) (81.6711) (-35.3563) (15.0337) (-36.3080) (81.5298) (30.1700)(26.4699)(34.5499)(34.9431) (73.4231) (-47.9399) (-44.8889) (75.7148) (62.0846) (-46.5421) (-48.4664) (61.6906) 0.0296*** 0.9242*** -0.0327***0.0434*** -0.0465*** 0.8966***0.1134***0.0957***0.1466***0.1486*** 0.7527*** -0.1127*** -0.0931*** 0.7865*** 0.8993 0.8682 0.7170 0.7469 0.6751*** -0.1348*** 0.6353 -0.1405*** 0.6749*** 0.6396 -statistics (in brackets) and adjusted R2’s of a pooled-panel regression. The standard errors to t nom i,t roe M,CF N nom i,t M,y bm N − M,e i,t e N 0.0086 0.9678*** 0.0100** 1.1659 -0.3574*** -0.3749** -0.2779 0.0379 0.3669** -0.4177*** 0.9189 0.0211 -0.0024 -0.0077 -0.0077 0.0005 0.0582 -0.0081 0.0044 0.0263*** -2.5812*** 0.1673 -0.3844** 0.1205 -0.2355 -1.4836*** 0.7026*** 0.0330 -0.0024 0.0024 0.0005 − 0.0265*** -0.0104*** 0.8897*** -0.1874 0.2027*** -0.6543*** -0.1741** -0.3214*** -0.9806*** 0.2512*** 0.7885 1 3 1 3 1 3 -statistics are clustered by firm-quarter (Petersen, 2009 ). The rows with the aggregate variables have been removed to save space. The t +1 +1 +1 +1 +1 +1 nom nom bmQ i,t bmQ i,t i,t volQ i,t volQ i,t i,t i,t meQ meQ i,t +1 M,e 1 1 1 1 1 1 i,t N M,y bm roe e − N M,CF +1 +1 +1 +1 +1 +1 nom nom nom nom nom nom t t t t t t Var/covar − N y y y y y y (book-to-market) (0.6870) (170.6688) (2.0682) (1.3641) (-2.6627) (-2.2308) (-1.1232) (0.1435) (2.3494) (-2.7007) (return-on-equity) (7.4453) (-5.5851) (89.3682) (-1.4739) (4.3727) (-8.4610) (-2.1843) (-2.9151) (-8.0636) (6.4250) Panel A: VAR coefficients (log excess return) (-0.6326) (0.8913) (4.7177) (-2.7604) (1.1695) (-2.1939) (0.5792) (-0.9378) (-8.1774) (3.9332) Panel B: News terms compute the dynamics of the aggregateempirical variables covariance matrix. are the Onand the same 99% right, as levels the in are correlations Table indicated (off-diagonal2.1 . with terms) *, Panel and ** B the and standard reports *** deviations descriptive respectively. (diagonal). statistics A Significance on at the the implied 90%, news 95% terms. On the left, the Table 2.8:Panel Firm A VAR model reports with the interaction OLS terms estimated coefficients, from February 1927 to October 2015. The columns with aggregate state variables, except for interest rate risk have also been removed to save space. CHAPTER 2. BETA 45 Real stock return Real dividends Excess stock return Real interest rate Real dividends Excess stock return Real interest rate Real dividends Excess stock return Real dividends* Excess stock return Real dividends* Excess stock return Real dividends* Excess stock return Real dividends* Real stock return Real dividends Volatility Dividend-to-price 1M T-bill rate (12m demeaned) Dividend-to-price Real interest rate Change 1M T-bill rate 10Y-2M term yield 1M T-bill rate (12m demeaned) Real interest rate Dividend yield Inflation rate Growth rate industrial production Book-to-market Profitability 10Y-3M term yield Shiller PE Value Spread 10Y-3M term yield Shiller PE Value Spread 10Y-3M term yield Shiller PE Value Spread Default Spread Stock return variance Shiller PE 3M T-bill rate Value Spread Default Spread StudyCampbell ( 1991 )Campbell and Ammer ( 1993 )Campbell and Mei ( 1993 ) 1952-1987 1927-1988 monthly DataVuolteenaho sample monthly ( 2002 ) Excess stock Frequency return Real stockCampbell return and State 1952-1987 Vuolteenaho variables ( 2004 ) monthlyCampbell, Polk, and Vuolteenaho 1929-2001 ( 2010 ) Excess stock return 1928-2001 1954-1996 monthlyCampbell, Giglio, and Polk ( 2013 ) annual Return annual components Excess stock return Excess stockCampbell Excess return stock et 1929-2010 return al. ( 2017 ) quarterly Excess stock return 1926-2011 quarterly Real stock return Table 2.9:Studies Literature marked overview with aggregate a VAR * models in the return components section assume that real interest rates are constant. 46
Table 2.10: Literature overview firm VAR models Studies marked with a * in the return components section assume that real interest rates are constant.
Study Data sample Frequency State variables Return components Vuolteenaho (2002) 1954-1996 annual Excess stock return Excess stock return Book-to-market Real dividends* Return-on-equity
Campbell, Polk and 1928-2001 annual Excess stock return Excess stock return Vuolteenaho (2010) Book-to-market Real dividends* Return-on-equity CHAPTER 2. BETA 47 = 12 2016 / 1 5 . 2006 risk premium interest rate cash flow 1996 1986 1976 1966 1956 1946 1936 1926 04 03 02 01 00 01 02 03 04 05 06 07 ...... 0 0 0 0 0 0 0 0 0 0 0 0
− − − − − − − 9439. The news terms are based on the aggregate VAR model displayed in Table . 2.1 fractions) (in return . Figure 2.1:This News figure terms plots through the trailing time exponentially weighted average quarterly news term in fractions. The smoothing parameter is set to 0 0 48 2011 risk premium interest rate cash flow 2001 1991 1981 1971 1961 1951 1941 9439. The prices are from specification 6 in Table 2.4 . 1931 . 40 30 20 10 10 20 30 00 ...... = 0 0 0 0 0 0 0 0 0 12 − − − /
1 5 fractions) (in beta of unit per return quarterly . Figure 2.2:This Prices through figure time plots0 the trailing exponentially weighted average quarterly prices of risk in fractions. The smoothing parameter is set to Chapter 3
Are Stock and Corporate Bond Markets Integrated? Evidence from Expected Returns
3.1 Introduction
Firms can finance themselves on public capital markets in two ways: first, by issuing equity, i.e. stocks, and second by issuing debt, i.e. corporate bonds. Since both instruments are claims on the same assets, their prices and expected returns should be linked to each other. As discussed below, several studies have therefore studied the integration of stock and corporate bond markets using various empirical methods. In this paper we add to this integration literature by directly comparing bond-implied expected equity returns to realized equity returns. For each firm, we construct the expected equity return that is implied by the credit spread on corporate bonds of this firm, following the approach of Campello, Chen, and Zhang (2008). We then analyze the cross-sectional relation between these bond-implied expected equity returns and average realized equity returns for the U.S. market. Surprisingly, we find a strong negative relation. Firms with high bond-implied expected equity returns have low average equity returns and vice versa. This suggests that corporate bonds and stocks are not priced consistently. The economic significance of this result is large. When sorting firms on the bond-implied expected equity return, we find a predicted equity return gap of 1.27% per month for the highest versus lowest decile portfolio, while the realized equity return gap equals -1.79% per month. 50
It is important to note that our analysis of expected and realized returns provides evidence for relative mispricing of corporate bonds versus stocks. We do not aim to explain the level of bond-implied expected returns, nor the level of realized average equity returns. For such an analysis, one would need to assume a specific asset pricing model and see if these levels are in line with the model predictions. In contrast, for our main result, the negative relation between expected bond and stock returns, we do not need to assume a specific asset pricing model. We do need to make a few other modeling assumptions. This concerns first of all the default probability, which is required to transform the credit spread to a corporate bond expected return. Second, our approach requires an estimate of the sensitivity of equity returns to corporate bond returns, which is needed to transform the corporate bond expected return to an expected equity return. Our benchmark approach uses a Merton (1974) model to estimate default probabilities, following Feldh¨utter and Schaefer (2016). We find the Merton model to be more adaptive than the hazard rate model of Campbell, Hilscher, and Szilagyi (2008) and better able to capture the strongly increasing probability of default for low-rated debt. To estimate equity-bond sensitivities, we employ a regression approach following Campello, Chen, and Zhang (2008). We perform a wide range of robustness checks on these two methods. We also perform robustness checks on the benchmark cross-sectional analysis, to check that results are stable over time and not concentrated in small, hard- to-arbitrage stocks. In addition, we correct for liquidity effects. We find that our key result survives all these robustness checks. We then proceed by trying to understand this lack of integration in more detail. First, we document that the anomalous realized equity returns to some extent reflect temporary mispricing in the equity market, but substantial mispricing remains even at longer horizons. Specifically, when we focus on the realized equity returns over a period of 5 years after sorting on the expected return, the realized return gap is -0.50% per month instead of -1.79%. Second, we analyze whether market risk or characteristics like size, book-to-market, momentum, profitability, and investments are priced differently in corporate bonds versus equities. We find some evidence that the market betas have a negative slope for average realized returns (in line with existing work on the “beta anomaly”; Frazzini and Pedersen, 2014), while the slope is positive for bond-implied returns. However, we find that these differences in pricing do not explain the negative relation between bond-implied and realized returns. In line with these results, the alpha of a portfolio that shorts stocks of firms with high bond-implied returns and buys stocks with low bond-implied returns is significantly positive, even when we control for the five factors in the Fama and French (2015) model, the Carhart (1997) momentum factor and the Quality- CHAPTER 3. STOCK-BOND INTEGRATION 51
Minus-Junk factor of Asness, Frazzini, and Pedersen (2017). Our paper relates to various streams in the literature. Most importantly, two recent studies also focus on the cross-sectional relation between expected returns in credit markets (corporate bonds and credit default swaps) and re- alized equity returns: Friewald, Wagner, and Zechner (2014) and Anginer and Yildizhan (2017). Both studies conclude that there is a positive relation be- tween the cross-section of credit risk premiums and average realized equity returns. Our results thus conflict with these studies. We now discuss both studies in more detail. Friewald, Wagner, and Zechner (2014) and our study differ in several as- pects. First, Friewald, Wagner, and Zechner (2014) use a different approach to construct credit risk premiums. Following Cochrane and Piazzesi (2005) they run predictive regressions of credit spread changes on forward credit spreads to obtain time series of the credit risk premium for each firm. Second, most of their analysis is in-sample, since they run the predictive regressions over the same period that is used to calculate realized equity returns. This is particularly important in this case. Consider a firm that has had excellent performance over the sample period with declining credit spreads and increas- ing equity prices. The in-sample estimation will then lead to high estimates for the credit risk premium as the credit spreads have compressed, while the average realized equity return will be high as well. Hence, the in-sample esti- mation biases towards finding a positive relation between credit risk premiums and equity returns.1 In contrast, our credit risk premiums are only based on information available at the given point in time and sorting stocks on these credit risk premiums thus delivers a tradable investment strategy. Third, our sample is much more extensive, covering on average 685 firms per month for a period of more than 20 years. Friewald, Wagner, and Zechner (2014) use credit default swaps to obtain credit spreads, and as a result their sample period is less than 10 years and includes just 491 unique firms. Anginer and Yildizhan (2017) use conceptually the same approach as we do to obtain the credit risk premium from credit spreads.2 However, they do not transform the credit risk premium to an expected, corporate bond-implied, equity return. Hence, they cannot perform the quantitative comparison of expected corporate bond-implied equity returns and realized equity returns. Moreover, the focus of their paper is to analyze how systematic default risk
1In their out-of-sample robustness check Friewald, Wagner, and Zechner (2014) find much weaker evidence for a positive relation between credit premiums and stock returns. 2Anginer and Yildizhan (2017) employ a hazard rate model to estimate the probability of default rather than the probability of default implied by the Merton (1974) model. In the robustness section we employ the hazard rate model of Campbell, Hilscher, and Szilagyi (2008) and obtain results that are qualitatively similar to our benchmark results. 52 betas relate to these expected and realized returns, while our main focus is to perform an extensive and quantitative analysis of the bond-implied equity returns and realized equity returns. Finally, Anginer and Yildizhan (2017) use a longer sample period (1980 to 2010) but have a smaller cross-section of firms: about 338 firms each month3, compared to 685 for our sample. Our work also relates to other studies on the integration of stock and cor- porate bond markets. In particular, a number of studies study integration by looking at the time-series relation between stock and corporate bond re- turns. Hence, these studies do not focus on pricing and expected returns, and our work thus complements this stream in the literature. Examples of stud- ies in this literature are Collin-Dufresne, Goldstein, and Martin (2001) and Demirovic, Guermat, and Tucker (2017), who focus on the contemporaneous relation between stock and bond returns, and conclude that stock and bond markets are not perfectly integrated.4 In addition, there is a large body of literature on lead-lag effects, concluding either that stock returns lead bond re- turns (Kwan, 1996; Gebhardt, Hvidkjaer, and Swaminathan, 2005; Downing, Underwood, and Xing, 2009; Haesen, Houweling, and Van Zundert, 2017) or vice versa (Bittlingmayer and Moser, 2014; Ben Dor and Xu, 2015), which is also indirect evidence of disintegration between the two markets as it suggests new information is not priced in in both assets at the same time.5 Note that, even if there is an imperfect time-series relation between stock and corporate bond returns, this does not necessarily imply that long-term expected returns are different, because this imperfect time-series relation might be caused by temporary illiquidity or price pressure effects, or by exposure to factors that are not priced. In addition, our work is related to several studies that focus on the pric- ing of equity market anomalies in bond markets (Chordia et al., 2017; Choi and Kim, 2017). If bond and stock markets are integrated, then well-known anomalies should be present in both stock and bond markets. Concretely, Choi and Kim (2017) study the cross-sectional pricing of known equity anomalies in the cross-section of corporate bond returns, and find that some anomalies are similarly priced (net issuance, gross profitability, idiosyncratic volatility, beta and accruals), but others are not (asset growth and momentum). Related
3Anginer and Yildizhan (2017) report 121,714 firm-month observations over the period January 1981 to December 2010, which means on average there are 338 firms in a month. 4Schaefer and Strebulaev (2008) find that, despite this imperfect time-series relation, the size of the exposure of corporate bonds to equity returns is similar to predictions of a Merton (1974) model. 5Relatedly, there are also studies focusing on the link between credit default swaps and stocks at the firm level, such as Duarte, Longstaff, and Yu (2007), Kapadia and Pu (2012), Hilscher, Pollet, and Wilson (2015), and Kiesel, Kolaric, and Schiereck (2016). CHAPTER 3. STOCK-BOND INTEGRATION 53 to Choi and Kim (2017) is the work of Chordia et al. (2017), who study the predictive power of profitability, asset growth, equity market capitalization, accruals and earnings surprises. They find that, after transaction costs, bonds are efficiently priced. In our analysis, we examine whether our results can be explained by a different effect of anomalies on stock and bond markets. We control for anomalies related to beta, size, book-to-market, momentum, prof- itability, investment, and quality-minus-junk, and find that our main result cannot be explained by a different presence of these anomalies in stock versus corporate bond markets. Finally, our paper is related to the literature that documents a relation between the default probability and average equity returns. They mostly document a negative relation: firms with a high default probability have low average returns. This result is often referred to as the “distress risk puzzle”. We show that our main result is not simply a restatement of this distress risk puzzle. First of all, we show that the relation between our bond-implied equity return and default probabilities is not monotonic. Second, we double-sort on default probabilities and the bond-implied equity return and continue to find evidence for a negative relation between bond-implied and realized returns. The remainder of this paper is organized as follows. In Section 3.2 we de- scribe the data. Section 3.3 describes how we obtain expected equity returns from corporate bond credit spreads. Section 3.4 presents the benchmark em- pirical results: we analyze the relation between the bond-implied expected equity returns and realized equity returns. Section 3.5 provides robustness checks and Section 3.6 concludes.
3.2 Data
For our empirical analyses, we use stock data from the Center of Research in Security Prices (CRSP) at a monthly frequency over the period January 1994 to December 2015. We only include common equity (share codes 10 or 11) and exclude financials (SIC codes 6000-6999) as their financial structure is very different from corporates (see also Campbell, Hilscher, and Szilagyi, 2008). To compute probabilities of default, we use accounting data from COM- PUSTAT Quarterly. The COMPUSTAT data is linked to CRSP using the CRSP/Compustat Merged database, and all accounting data is lagged for two months to account for the reporting lag. For the bond data we use monthly constituent data of the Bloomberg Bar- clays U.S. Corporate Investment Grade and Bloomberg Barclays U.S. Cor- porate High Yield indices, formerly known as the Lehman Brothers Fixed 54
Income database. This dataset includes per bond the (option adjusted) spread duration, which is adjusted for embedded options by Bloomberg. This is nec- essary in case bonds are likely to be called ahead of their maturity date, as the standard modified duration would overstate the interest rate sensitivity. Using the option-adjusted spread duration, the bond is matched with the ap- propriate U.S. Treasury bond. The credit spread (excess return) is computed as the difference between the yield (return) on the corporate bond and the yield (return) on the duration-matched Treasury bond. As the embedded option-adjustment for the spread duration and the credit spread fields are not available prior to January 1994, we start at this date. Furthermore, the credit rating is computed as the middle rating of S&P, Moody’s and Fitch if all three are available, or the worst rating if only two are available, in line with the Bloomberg-Barclays index methodology. We link the bond data to the stock/accounting data using CUSIP’s and if not possible, hand-match the data, while taking M&A activity into account. See Appendix 3.A for details. A single stock can have multiple bonds associ- ated. To make the selected bonds as comparable as possible between stocks, we always pick a senior unsecured bond. If multiple senior unsecured bonds exist, we pick the one with the spread duration closest to 5 years in order to reduce the dispersion across maturities. Finally, we obtain the 1-month T-bill rate, the five Fama and French (2015) factors and the Carhart (1997) Momentum factor return series from the web- site of Kenneth French.6 The Quality-Minus-Junk factor is from the website of AQR.7
3.3 Methodology
Our empirical framework is based on the idea that the equity and bond risk premium of a firm are linked to each other as both are contingent claims on the same firm assets (Merton, 1974). In the Merton (1974) model, the bond and equity value are dependent on the firm value, the interest rate and the asset volatility. To arrive at a parsimonious model, we assume, following Campello, Chen, and Zhang (2008)8, the interest rate term structure to be flat and deterministic, and the asset volatility to be at most a function of asset value. This leaves the firm value as the sole driver of expected bond and equity returns.
6http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data library.html 7https://www.aqr.com/library/data-sets/ 8Campello, Chen, and Zhang (2008) also include a convexity effect. As we estimate the expected return to maturity, convexity effects are not relevant in our setting. CHAPTER 3. STOCK-BOND INTEGRATION 55
3.3.1 Obtaining expected equity returns
In this framework, the expected return of the equity of firm j can be written as a function of the expected return on the bond:
δE/E rE = y + (rB) − y (3.1) Et j g,t δB/B Et j g,t
9 δE/E where yg,t is the yield of the duration-matched Treasury , δB/B the elasticity B of equity returns to bond returns and Et(rj ) − yg,t the expected return of the bond of firm j over the duration matched Treasury bond. To implement this equation, we need to have estimates for the elasticity and the expected bond return. We follow Campello, Chen, and Zhang (2008), Bongaerts, De Jong, and Driessen (2017), and others for computing the expected bond return over Trea- suries. Specifically, we assume defaults occur at maturity only and approxi- mate the coupon-paying bonds by zero-coupon bonds with maturity equal to the duration of the original bonds. Then, an investment of $1 in the zero- Tj,t coupon corporate bond pays off (1 + yg,t + sj,t) (1 − Lπj,t) by expectation at maturity Tj,t. Here, sj,t is the credit spread, πj,t the cumulative probability of default to maturity, L is the loss rate in case of default, and Tj,t the time to maturity, which we set equal to the duration of the bond. It then follows that the annualized expected bond return in excess of a maturity-matched Treasury bond is given by